すべてのモデルのランキング(記録)

出典区分や測定条件(比較グループ)の違う値は混ぜずに順位を付けます。総合は同じ比較グループの集合に値のあるモデル同士だけで集計し、コストはベンチマークをまたいで合算しません。方法の詳細はこのサイトについてにあります。

旧モデルと Claude Code / Codex で選べないモデルを含む、掲載しているすべてのモデル(57 モデル)で順位・百分位・前線を付けた記録です。今選べるモデルだけの順位はランキングにあります(順位を付ける顔ぶれが違うので、順位も違います)。

総合(性能優先)

比較グループの集合が同じ(その集合のすべてに値がある)model × effort 同士だけで集計します。各グループの中で集計対象だけに順位を付けて百分位(1 位 = 100、最下位 = 0)にし、同じベンチマークのグループどうしで平均してから、ベンチマーク間の平均で並べます(同じベンチマークの比較グループが複数あっても重く数えません)。集計ごとに対象のグループが違うので、別の集計の数値どうしは比べられません。

集計はモデル数 → ベンチマーク数 → 件数の多い順に並べ、今 Claude Code / Codex で選べるモデルを最も多く含む集計を先頭で開きます。残りは「ほかの集計」に出典区分ごとに分けて畳んでいます。

集計の条件: モデル 3 つ以上(effort 違いは 1 モデル)とベンチマーク 3 つ以上(同じベンチマークの複数の比較グループは 1 と数えます)。カバレッジ n/N はベンチマーク単位で、N は同じ出典区分で値のあるベンチマークの数です。各グループの「計測・報告」は、その集計に入った値の計測・報告と件数です(公式発表でも、他社のモデルの値を発表元が計測していることがあります)。

第三者・36 モデル・36 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(36)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(36)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(36)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)
1GPT-6 AstraMax959/1495.8%93.7%97.6%
2GPT-6.1 SolMax949/1495.4%93.7%100.0%
3Gemini 4 CarbonHigh928/1496.5%90.5%95.1%
4Claude Fable 5.2Max886/1493.9%95.1%100.0%
5GPT-6.1 TerraMax846/1494.3%90.9%92.7%
6Claude Sonnet 5.5Max817/1495.6%88.8%80.5%
7GPT-6 SolMax807/1494.3%89.8%90.0%
8Gemini 4 ArgonHigh809/1496.4%86.0%78.1%
9GPT-5.6 SolMax736/1493.5%89.1%82.9%
10Claude Opus 5.5Max707/1490.6%91.2%95.0%
11GPT-5.6 TerraMax667/1493.3%86.0%70.7%
12Kimi K3.1 ProMax666/1494.2%78.3%65.9%
13Gemini 3.7 FlashHigh577/1494.8%71.6%36.6%
14GPT-5.6 LunaMax567/1491.6%82.1%61.0%
15Claude Fable 5Max557/1485.9%87.0%90.2%
16Claude Opus 4.8Max538/1491.0%80.0%56.1%
17DeepSeek-V4.1-Flash-1017Max506/1493.2%71.2%46.3%
18Qwen3.8-MaxXhigh508/1492.7%74.7%46.3%
19GPT-6 LunaMax499/1490.5%79.0%56.1%
20Kimi K3Max498/1493.1%72.2%39.0%
21Gemini 3.8 FlashHigh467/1495.4%68.4%22.0%
22Claude Haiku 5.5Max426/1489.7%75.1%46.3%
23Grok 4.6Xhigh407/1493.2%66.0%31.7%
24Gemini 3.1 Pro Preview記載なし385/1494.1%59.7%26.8%
25GLM-5.3Max356/1490.9%68.8%29.3%
26Gemini 3.6 FlashHigh357/1494.1%59.0%22.0%
27DeepSeek-V4-Pro-0813Max307/1491.7%64.6%26.8%
28Claude Opus 4.7Max306/1486.4%70.2%31.7%
29DeepSeek-V4.1-Flash記載なし278/1489.8%67.4%26.8%
30Claude Opus 4.6Max238/1488.4%66.0%26.8%
31Grok 4.7Xhigh227/1492.7%53.0%17.1%
32Mistral Large 4High187/1488.1%58.6%26.8%
33GLM-5.3-FlashMax155/1490.2%55.8%17.1%
34Gemini 3 Flash Preview記載なし55/1483.2%51.2%17.1%
35Gemini 3.5 Flash-LiteHigh25/1483.3%26.0%0.0%
35Gemini 2.5 Pro記載なし25/1485.3%24.6%0.0%
ほかの集計(74 件)

公式

公式・5 モデル・5 件 × 6 グループ(カバレッジ 6/90 ベンチマーク)FrontierSWE 2、AutomationBench、Terminal-Bench-Science 0.1、OSWorld 2.1(offline set・partial)、AA-Briefcase v1.1(rubric)、AA-Briefcase v1.1(Elo)
  1. FrontierSWE 2公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(3)・記載なし(2)
  2. AutomationBench公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(3)・記載なし(2)
  3. Terminal-Bench-Science 0.1公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(3)・記載なし(2)
  4. OSWorld 2.1(offline set・partial)公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(3)・記載なし(2)
  5. AA-Briefcase v1.1(rubric)公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(3)・記載なし(2)
  6. AA-Briefcase v1.1(Elo)公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(3)・記載なし(2)
順位モデルeffort平均百分位値のあるベンチマーク1. FrontierSWE 22. AutomationBench3. Terminal-Bench-Science 0.14. OSWorld 2.1(offline set・partial)5. AA-Briefcase v1.1(rubric)6. AA-Briefcase v1.1(Elo)
1Claude Fable 5.2記載なし927/9069.9%51.4%74.6%89.3%66.1%1,785
2Claude Opus 5.5記載なし5414/9062.3%42.5%58.7%87.2%61.2%1,800
3Gemini 4 Carbon記載なし507/9061.2%53.9%64.8%81.5%61.8%1,611
4GPT-6 Astra記載なし2912/9065.5%41.4%68.1%79.5%52.0%1,569
5Claude Fable 5.1記載なし2511/9056.3%31.4%52.6%83.5%61.5%1,669
公式・5 モデル・25 件 × 3 グループ(カバレッジ 3/90 ベンチマーク)Terminal-Bench 4.0、FrontierCode v1.1(main set)、GDPval-AA v2.1
  1. Terminal-Bench 4.0公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(15)・OpenAI(10)
  2. FrontierCode v1.1(main set)公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(15)・記載なし(10)
  3. GDPval-AA v2.1公式・Introducing Claude Opus 5.5計測・報告: Anthropic(15)・記載なし(10)
順位モデルeffort平均百分位値のあるベンチマーク1. Terminal-Bench 4.02. FrontierCode v1.1(main set)3. GDPval-AA v2.1
1Claude Opus 5.5Max9716/9064.8%54.4%1,846
2Claude Opus 5.5Xhigh9013/9066.4%51.4%1,820
3Claude Opus 5.5High8813/9064.2%54.0%1,692
4Claude Opus 5.5Med8013/9057.6%54.6%1,576
5Claude Fable 5.1Max7310/9055.8%50.3%1,735
6GPT-6 AstraMax695/9056.7%53.3%1,542
7GPT-6 AstraHigh655/9057.9%50.9%1,485
8Claude Fable 5.1Xhigh6410/9051.3%48.7%1,721
9Claude Opus 5Max626/9052.3%48.0%1,708
9GPT-6 AstraXhigh625/9057.6%50.6%1,516
11Claude Fable 5.1High5810/9049.4%50.3%1,617
12Claude Fable 5.1Med5110/9043.4%50.9%1,536
13Claude Opus 5Med506/9041.2%53.4%1,476
14Claude Opus 5High486/9047.0%48.0%1,581
15Claude Opus 5Xhigh476/9050.6%43.6%1,676
15GPT-6 AstraMed475/9053.9%48.8%1,468
17Claude Fable 5.1Low4310/9040.2%52.8%1,450
18GPT-5.6 SolMax405/9037.3%47.5%1,588
19GPT-5.6 SolXhigh315/9028.5%46.8%1,548
20GPT-6 AstraLow285/9049.7%45.3%1,366
21GPT-5.6 SolHigh195/9026.1%45.1%1,480
22Claude Opus 5.5Low1813/9038.5%47.3%1,224
23Claude Opus 5Low106/9028.5%42.0%1,294
24GPT-5.6 SolMed85/9020.9%39.9%1,403
25GPT-5.6 SolLow15/907.9%35.4%1,289
公式・4 モデル・4 件 × 7 グループ(カバレッジ 7/90 ベンチマーク)FrontierSWE 2、AutomationBench、Terminal-Bench-Science 0.1、OSWorld 2.1(offline set・partial)、RoboDojo Sim、AA-Briefcase v1.1(rubric)、AA-Briefcase v1.1(Elo)
  1. FrontierSWE 2公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(3)・記載なし(1)
  2. AutomationBench公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(3)・記載なし(1)
  3. Terminal-Bench-Science 0.1公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(3)・記載なし(1)
  4. OSWorld 2.1(offline set・partial)公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(3)・記載なし(1)
  5. RoboDojo Sim公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(4)
  6. AA-Briefcase v1.1(rubric)公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(3)・記載なし(1)
  7. AA-Briefcase v1.1(Elo)公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(3)・記載なし(1)
順位モデルeffort平均百分位値のあるベンチマーク1. FrontierSWE 22. AutomationBench3. Terminal-Bench-Science 0.14. OSWorld 2.1(offline set・partial)5. RoboDojo Sim6. AA-Briefcase v1.1(rubric)7. AA-Briefcase v1.1(Elo)
1Claude Fable 5.2記載なし957/9069.9%51.4%74.6%89.3%64.9%66.1%1,785
2Claude Opus 5.5記載なし5714/9062.3%42.5%58.7%87.2%39.4%61.2%1,800
3GPT-6 Astra記載なし2912/9065.5%41.4%68.1%79.5%29.8%52.0%1,569
4Claude Fable 5.1記載なし1911/9056.3%31.4%52.6%83.5%27.6%61.5%1,669
公式・4 モデル・20 件 × 4 グループ(カバレッジ 4/90 ベンチマーク)Terminal-Bench 4.0、FrontierCode v1.1(main set)、CursorBench 4.0、GDPval-AA v2.1
  1. Terminal-Bench 4.0公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(15)・OpenAI(5)
  2. FrontierCode v1.1(main set)公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(15)・記載なし(5)
  3. CursorBench 4.0公式・Introducing Claude Opus 5.5計測・報告: Anthropic(15)・記載なし(5)
  4. GDPval-AA v2.1公式・Introducing Claude Opus 5.5計測・報告: Anthropic(15)・記載なし(5)
順位モデルeffort平均百分位値のあるベンチマーク1. Terminal-Bench 4.02. FrontierCode v1.1(main set)3. CursorBench 4.04. GDPval-AA v2.1
1Claude Opus 5.5Max9716/9064.8%54.4%57.8%1,846
2Claude Opus 5.5Xhigh9013/9066.4%51.4%56.0%1,820
3Claude Opus 5.5High8613/9064.2%54.0%56.0%1,692
4Claude Opus 5.5Med7913/9057.6%54.6%52.5%1,576
5Claude Fable 5.1Max7710/9055.8%50.3%51.8%1,735
6Claude Fable 5.1Xhigh7010/9051.3%48.7%51.6%1,721
7Claude Opus 5Max646/9052.3%48.0%46.6%1,708
8Claude Fable 5.1High6310/9049.4%50.3%49.2%1,617
9Claude Fable 5.1Med5410/9043.4%50.9%46.8%1,536
10Claude Opus 5Xhigh506/9050.6%43.6%46.1%1,676
11Claude Opus 5High486/9047.0%48.0%44.7%1,581
12Claude Fable 5.1Low4610/9040.2%52.8%45.1%1,450
12Claude Opus 5Med466/9041.2%53.4%43.3%1,476
14GPT-5.6 SolMax375/9037.3%47.5%41.7%1,588
15GPT-5.6 SolXhigh265/9028.5%46.8%37.7%1,548
16Claude Opus 5.5Low2513/9038.5%47.3%43.7%1,224
17GPT-5.6 SolHigh185/9026.1%45.1%35.7%1,480
18Claude Opus 5Low156/9028.5%42.0%40.7%1,294
19GPT-5.6 SolMed85/9020.9%39.9%31.1%1,403
20GPT-5.6 SolLow15/907.9%35.4%24.6%1,289
公式・4 モデル・20 件 × 4 グループ(カバレッジ 4/90 ベンチマーク)Terminal-Bench 4.0、FrontierCode v1.1(main set)、GDPval-AA v2.1、AutomationBench
  1. Terminal-Bench 4.0公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(10)・OpenAI(10)
  2. FrontierCode v1.1(main set)公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(10)・記載なし(10)
  3. GDPval-AA v2.1公式・Introducing Claude Opus 5.5計測・報告: Anthropic(10)・記載なし(10)
  4. AutomationBench公式・Introducing Claude Opus 5.5計測・報告: Zapier(20)
順位モデルeffort平均百分位値のあるベンチマーク1. Terminal-Bench 4.02. FrontierCode v1.1(main set)3. GDPval-AA v2.14. AutomationBench
1Claude Opus 5.5Max9616/9064.8%54.4%1,84640.0%
2Claude Opus 5.5Xhigh8713/9066.4%51.4%1,82034.4%
3Claude Opus 5.5High8313/9064.2%54.0%1,69232.0%
4GPT-6 AstraMax755/9056.7%53.3%1,54241.4%
5Claude Opus 5.5Med7313/9057.6%54.6%1,57628.6%
6GPT-6 AstraHigh705/9057.9%50.9%1,48537.1%
7GPT-6 AstraXhigh695/9057.6%50.6%1,51639.0%
8Claude Opus 5Max616/9052.3%48.0%1,70826.9%
9GPT-6 AstraMed555/9053.9%48.8%1,46834.1%
10GPT-5.6 SolMax505/9037.3%47.5%1,58828.8%
11Claude Opus 5Xhigh466/9050.6%43.6%1,67625.3%
12Claude Opus 5Med456/9041.2%53.4%1,47623.9%
13Claude Opus 5High446/9047.0%48.0%1,58120.6%
14GPT-6 AstraLow385/9049.7%45.3%1,36630.3%
15GPT-5.6 SolXhigh385/9028.5%46.8%1,54826.3%
16GPT-5.6 SolHigh255/9026.1%45.1%1,48024.8%
17Claude Opus 5.5Low2213/9038.5%47.3%1,22423.3%
18Claude Opus 5Low136/9028.5%42.0%1,29420.4%
19GPT-5.6 SolMed95/9020.9%39.9%1,40319.6%
20GPT-5.6 SolLow15/907.9%35.4%1,28911.7%
公式・4 モデル・20 件 × 3 グループ(カバレッジ 3/90 ベンチマーク)Terminal-Bench 4.0、FrontierCode v1.1(main set)、CursorBench 4.0
  1. Terminal-Bench 4.0公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(15)・OpenAI(5)
  2. FrontierCode v1.1(main set)公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(15)・記載なし(5)
  3. CursorBench 4.0公式・Introducing Claude Sonnet 5.5計測・報告: Anthropic(15)・記載なし(5)
順位モデルeffort平均百分位値のあるベンチマーク1. Terminal-Bench 4.02. FrontierCode v1.1(main set)3. CursorBench 4.0
1Claude Opus 5.5Max9516/9064.8%54.4%57.8%
2Claude Opus 5.5High8913/9064.2%54.0%56.0%
2Claude Opus 5.5Xhigh8913/9066.4%51.4%56.0%
4Claude Opus 5.5Med8213/9057.6%54.6%52.5%
5Claude Sonnet 5.5Xhigh817/9061.5%52.1%53.1%
6Claude Sonnet 5.5Max797/9070.6%46.2%55.5%
7Claude Sonnet 5.5High707/9043.0%49.4%47.8%
8Claude Opus 5.5Low6313/9038.5%47.3%43.7%
9GPT-5.6 SolMax615/9037.3%47.5%41.7%
10GPT-5.6 SolXhigh515/9028.5%46.8%37.7%
11Claude Sonnet 5.5Med427/9028.8%36.5%39.2%
11GPT-5.6 SolHigh425/9026.1%45.1%35.7%
13Claude Sonnet 5Max324/9010.3%42.4%34.1%
14GPT-5.6 SolMed305/9020.9%39.9%31.1%
15Claude Sonnet 5Xhigh284/907.0%42.7%32.0%
16Claude Sonnet 5.5Low267/9020.0%29.3%35.8%
17Claude Sonnet 5High184/904.5%39.4%30.8%
18GPT-5.6 SolLow145/907.9%35.4%24.6%
19Claude Sonnet 5Med94/904.4%35.2%28.0%
20Claude Sonnet 5Low04/903.2%28.7%24.1%
公式・3 モデル・3 件 × 8 グループ(カバレッジ 8/90 ベンチマーク)FrontierSWE 2、AutomationBench、Terminal-Bench-Science 0.1、OSWorld 2.1(offline set・partial)、RoboDojo Sim、AA-Briefcase v1.1(rubric)、AA-Briefcase v1.1(Elo)、Remote Labor Index(RLI)
  1. FrontierSWE 2公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(2)・記載なし(1)
  2. AutomationBench公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(2)・記載なし(1)
  3. Terminal-Bench-Science 0.1公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(2)・記載なし(1)
  4. OSWorld 2.1(offline set・partial)公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(2)・記載なし(1)
  5. RoboDojo Sim公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(3)
  6. AA-Briefcase v1.1(rubric)公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(2)・記載なし(1)
  7. AA-Briefcase v1.1(Elo)公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(2)・記載なし(1)
  8. Remote Labor Index(RLI)公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(2)・記載なし(1)
順位モデルeffort平均百分位値のあるベンチマーク1. FrontierSWE 22. AutomationBench3. Terminal-Bench-Science 0.14. OSWorld 2.1(offline set・partial)5. RoboDojo Sim6. AA-Briefcase v1.1(rubric)7. AA-Briefcase v1.1(Elo)8. Remote Labor Index(RLI)
1Claude Opus 5.5記載なし8114/9062.3%42.5%58.7%87.2%39.4%61.2%1,80021.3%
2GPT-6 Astra記載なし4412/9065.5%41.4%68.1%79.5%29.8%52.0%1,56920.8%
3Claude Fable 5.1記載なし2511/9056.3%31.4%52.6%83.5%27.6%61.5%1,66917.9%
公式・3 モデル・3 件 × 6 グループ(カバレッジ 6/90 ベンチマーク)SWE-Bench Pro、OSWorld-Verified、CharXiv Reasoning、MMMU-Pro、MRCR v2(8-needle・128k)、MRCR v2(8-needle・1M)
  1. SWE-Bench Pro公式・Gemini 3.5 Flash model card計測・報告: Google(3)
  2. OSWorld-Verified公式・Gemini 3.5 Flash model card計測・報告: Google(3)
  3. CharXiv Reasoning公式・Gemini 3.5 Flash model card計測・報告: Google(3)
  4. MMMU-Pro公式・Gemini 3.5 Flash model card計測・報告: Google(3)
  5. MRCR v2(8-needle・128k)公式・Gemini 3.5 Flash model card計測・報告: Google(3)
  6. MRCR v2(8-needle・1M)公式・Gemini 3.5 Flash model card計測・報告: Google(3)
順位モデルeffort平均百分位値のあるベンチマーク1. SWE-Bench Pro2. OSWorld-Verified3. CharXiv Reasoning4. MMMU-Pro5. MRCR v2(8-needle・128k)6. MRCR v2(8-needle・1M)
1Gemini 3.5 Flash記載なし927/9055.1%78.4%84.2%83.6%77.3%26.6%
2Gemini 3.1 Pro Preview記載なし506/9054.2%76.2%83.3%80.5%84.9%26.3%
3Gemini 3 Flash Preview記載なし87/9049.6%65.1%80.3%81.2%67.2%22.1%
公式・3 モデル・15 件 × 5 グループ(カバレッジ 5/90 ベンチマーク)Terminal-Bench 4.0、FrontierCode v1.1(main set)、CursorBench 4.0、GDPval-AA v2.1、AutomationBench
  1. Terminal-Bench 4.0公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(10)・OpenAI(5)
  2. FrontierCode v1.1(main set)公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(10)・記載なし(5)
  3. CursorBench 4.0公式・Introducing Claude Opus 5.5計測・報告: Anthropic(10)・記載なし(5)
  4. GDPval-AA v2.1公式・Introducing Claude Opus 5.5計測・報告: Anthropic(10)・記載なし(5)
  5. AutomationBench公式・Introducing Claude Opus 5.5計測・報告: Zapier(15)
順位モデルeffort平均百分位値のあるベンチマーク1. Terminal-Bench 4.02. FrontierCode v1.1(main set)3. CursorBench 4.04. GDPval-AA v2.15. AutomationBench
1Claude Opus 5.5Max9716/9064.8%54.4%57.8%1,84640.0%
2Claude Opus 5.5Xhigh8913/9066.4%51.4%56.0%1,82034.4%
3Claude Opus 5.5High8513/9064.2%54.0%56.0%1,69232.0%
4Claude Opus 5.5Med7613/9057.6%54.6%52.5%1,57628.6%
5Claude Opus 5Max716/9052.3%48.0%46.6%1,70826.9%
6Claude Opus 5Xhigh546/9050.6%43.6%46.1%1,67625.3%
7GPT-5.6 SolMax535/9037.3%47.5%41.7%1,58828.8%
8Claude Opus 5High516/9047.0%48.0%44.7%1,58120.6%
9Claude Opus 5Med476/9041.2%53.4%43.3%1,47623.9%
10GPT-5.6 SolXhigh365/9028.5%46.8%37.7%1,54826.3%
11Claude Opus 5.5Low3313/9038.5%47.3%43.7%1,22423.3%
12GPT-5.6 SolHigh275/9026.1%45.1%35.7%1,48024.8%
13Claude Opus 5Low196/9028.5%42.0%40.7%1,29420.4%
14GPT-5.6 SolMed105/9020.9%39.9%31.1%1,40319.6%
15GPT-5.6 SolLow15/907.9%35.4%24.6%1,28911.7%
公式・3 モデル・15 件 × 5 グループ(カバレッジ 5/90 ベンチマーク)Terminal-Bench 4.0、FrontierCode v1.1(main set)、CursorBench 4.0、GDPval-AA v2.1、WANDR
  1. Terminal-Bench 4.0公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(15)
  2. FrontierCode v1.1(main set)公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(15)
  3. CursorBench 4.0公式・Introducing Claude Opus 5.5計測・報告: Anthropic(15)
  4. GDPval-AA v2.1公式・Introducing Claude Opus 5.5計測・報告: Anthropic(15)
  5. WANDR公式・Introducing Claude Opus 5.5計測・報告: Anthropic(15)
順位モデルeffort平均百分位値のあるベンチマーク1. Terminal-Bench 4.02. FrontierCode v1.1(main set)3. CursorBench 4.04. GDPval-AA v2.15. WANDR
1Claude Opus 5.5Max9716/9064.8%54.4%57.8%1,84672.3%
2Claude Opus 5.5Xhigh8813/9066.4%51.4%56.0%1,82071.3%
3Claude Opus 5.5High7913/9064.2%54.0%56.0%1,69267.3%
4Claude Fable 5.1Max7210/9055.8%50.3%51.8%1,73568.7%
5Claude Opus 5.5Med6313/9057.6%54.6%52.5%1,57662.8%
5Claude Fable 5.1Xhigh6310/9051.3%48.7%51.6%1,72167.7%
7Claude Opus 5Max546/9052.3%48.0%46.6%1,70867.2%
8Claude Fable 5.1High4910/9049.4%50.3%49.2%1,61766.7%
9Claude Opus 5Xhigh416/9050.6%43.6%46.1%1,67667.0%
10Claude Fable 5.1Med4010/9043.4%50.9%46.8%1,53664.5%
11Claude Opus 5High346/9047.0%48.0%44.7%1,58164.6%
12Claude Fable 5.1Low3110/9040.2%52.8%45.1%1,45063.3%
13Claude Opus 5Med296/9041.2%53.4%43.3%1,47658.1%
14Claude Opus 5.5Low713/9038.5%47.3%43.7%1,22431.2%
15Claude Opus 5Low36/9028.5%42.0%40.7%1,29450.5%
公式・3 モデル・15 件 × 5 グループ(カバレッジ 5/90 ベンチマーク)FrontierCode 2.0、WANDR(relaxed score)、BenchCAD 2.1(ツールあり)、Humanity's Last Exam(ツールあり)、Chartography(ツールなし)
  1. FrontierCode 2.0公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(15)
  2. WANDR(relaxed score)公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(15)
  3. BenchCAD 2.1(ツールあり)公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(15)
  4. Humanity's Last Exam(ツールあり)公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(15)
  5. Chartography(ツールなし)公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(15)
順位モデルeffort平均百分位値のあるベンチマーク1. FrontierCode 2.02. WANDR(relaxed score)3. BenchCAD 2.1(ツールあり)4. Humanity's Last Exam(ツールあり)5. Chartography(ツールなし)
1Claude Fable 5.2Max995/9045.6%82.1%83.574.2%73.4%
2Claude Fable 5.2Xhigh945/9049.1%80.6%74.173.4%71.4%
3Claude Fable 5.2High836/9045.3%79.3%6171.5%69.8%
4Claude Opus 5.5Max7616/9045.5%72.3%65.367.7%64.4%
5Claude Fable 5.2Med705/9037.6%73.6%32.468.6%66.4%
6Claude Opus 5.5Xhigh6313/9043.4%71.3%54.566.4%62.4%
7Claude Fable 5.2Low575/9034.2%71.4%26.968.4%64.2%
8Claude Opus 5.5High4413/9034.5%67.3%29.763.9%59.2%
9Claude Fable 5.1Xhigh4310/9032.4%67.7%40.565.2%47.0%
9Claude Fable 5.1Max4310/9031.7%68.7%44.465.6%44.8%
11Claude Fable 5.1High2610/9025.7%66.7%25.863.1%46.0%
12Claude Opus 5.5Med2413/9030.1%62.8%19.163.0%59.0%
13Claude Fable 5.1Med1410/9021.5%64.5%15.762.4%42.2%
14Claude Opus 5.5Low713/9014.7%31.2%6.357.4%57.0%
14Claude Fable 5.1Low710/9020.6%63.3%12.160.2%36.6%
公式・3 モデル・15 件 × 4 グループ(カバレッジ 4/90 ベンチマーク)Terminal-Bench 4.0、FrontierCode v1.1(main set)、GDPval-AA v2.1、FrontierCode 2.0
  1. Terminal-Bench 4.0公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(10)・OpenAI(5)
  2. FrontierCode v1.1(main set)公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(10)・記載なし(5)
  3. GDPval-AA v2.1公式・Introducing Claude Opus 5.5計測・報告: Anthropic(10)・記載なし(5)
  4. FrontierCode 2.0公式・Introducing Claude Fable 5.2 and Claude Mythos 5.2計測・報告: Anthropic(10)・記載なし(5)
順位モデルeffort平均百分位値のあるベンチマーク1. Terminal-Bench 4.02. FrontierCode v1.1(main set)3. GDPval-AA v2.14. FrontierCode 2.0
1Claude Opus 5.5Max9616/9064.8%54.4%1,84645.5%
2Claude Opus 5.5Xhigh8813/9066.4%51.4%1,82043.4%
3Claude Opus 5.5High7513/9064.2%54.0%1,69234.5%
4GPT-6 AstraMax685/9056.7%53.3%1,54242.8%
5Claude Opus 5.5Med6313/9057.6%54.6%1,57630.1%
6GPT-6 AstraHigh585/9057.9%50.9%1,48538.3%
7GPT-6 AstraXhigh565/9057.6%50.6%1,51641.4%
8Claude Fable 5.1Max5310/9055.8%50.3%1,73531.7%
9Claude Fable 5.1Xhigh4510/9051.3%48.7%1,72132.4%
10GPT-6 AstraMed385/9053.9%48.8%1,46836.8%
11Claude Fable 5.1High3510/9049.4%50.3%1,61725.7%
12Claude Fable 5.1Med3110/9043.4%50.9%1,53621.5%
13Claude Fable 5.1Low2510/9040.2%52.8%1,45020.6%
14GPT-6 AstraLow185/9049.7%45.3%1,36631.1%
15Claude Opus 5.5Low213/9038.5%47.3%1,22414.7%
公式・3 モデル・15 件 × 4 グループ(カバレッジ 4/90 ベンチマーク)Terminal-Bench 4.0、FrontierCode v1.1(main set)、CursorBench 4.0、AA-Briefcase v1.1(Elo)
  1. Terminal-Bench 4.0公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(15)
  2. FrontierCode v1.1(main set)公式・Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ)計測・報告: Anthropic(15)
  3. CursorBench 4.0公式・Introducing Claude Sonnet 5.5計測・報告: Anthropic(15)
  4. AA-Briefcase v1.1(Elo)公式・Introducing Claude Sonnet 5.5計測・報告: Anthropic(15)
順位モデルeffort平均百分位値のあるベンチマーク1. Terminal-Bench 4.02. FrontierCode v1.1(main set)3. CursorBench 4.04. AA-Briefcase v1.1(Elo)
1Claude Opus 5.5Max9516/9064.8%54.4%57.8%1,822
2Claude Opus 5.5Xhigh8513/9066.4%51.4%56.0%1,780
3Claude Opus 5.5High8113/9064.2%54.0%56.0%1,705
4Claude Sonnet 5.5Max807/9070.6%46.2%55.5%1,811
5Claude Sonnet 5.5Xhigh757/9061.5%52.1%53.1%1,746
6Claude Opus 5.5Med7313/9057.6%54.6%52.5%1,642
7Claude Sonnet 5.5High597/9043.0%49.4%47.8%1,634
8Claude Opus 5.5Low4813/9038.5%47.3%43.7%1,285
9Claude Sonnet 5.5Med397/9028.8%36.5%39.2%1,461
10Claude Sonnet 5Max344/9010.3%42.4%34.1%1,359
11Claude Sonnet 5Xhigh294/907.0%42.7%32.0%1,274
12Claude Sonnet 5.5Low257/9020.0%29.3%35.8%1,264
13Claude Sonnet 5High184/904.5%39.4%30.8%1,177
14Claude Sonnet 5Med94/904.4%35.2%28.0%1,056
15Claude Sonnet 5Low04/903.2%28.7%24.1%923
公式・3 モデル・11 件 × 4 グループ(カバレッジ 4/90 ベンチマーク)OSWorld 2.1(offline subset)、GDPval-AA v2.1、Humanity's Last Exam(ツールなし)、Terminal-Bench 4.0
  1. OSWorld 2.1(offline subset)公式・Introducing Claude Haiku 5.5計測・報告: Anthropic(11)
  2. GDPval-AA v2.1公式・Introducing Claude Haiku 5.5計測・報告: Anthropic(11)
  3. Humanity's Last Exam(ツールなし)公式・Introducing Claude Haiku 5.5計測・報告: Anthropic(11)
  4. Terminal-Bench 4.0公式・Introducing Claude Haiku 5.5計測・報告: Anthropic(11)
順位モデルeffort平均百分位値のあるベンチマーク1. OSWorld 2.1(offline subset)2. GDPval-AA v2.13. Humanity's Last Exam(ツールなし)4. Terminal-Bench 4.0
1Claude Sonnet 5.5Max1007/9083.9%1,84056.9%70.6%
2Claude Sonnet 5.5Xhigh907/9081.1%1,73153.0%61.5%
3Claude Sonnet 5.5High787/9073.2%1,55147.9%43.0%
4Claude Haiku 5.5Max734/9072.4%1,62045.9%39.2%
5Claude Haiku 5.5Xhigh604/9067.6%1,51344.1%31.5%
6Claude Sonnet 5.5Med487/9066.0%1,32443.2%28.8%
7Claude Haiku 5.5High404/9061.3%1,42039.7%24.8%
8Claude Sonnet 5.5Low287/9057.9%1,17941.4%20.0%
9Claude Haiku 5.5Med254/9053.3%1,27735.5%20.3%
10Claude Haiku 5.5Low104/9042.0%1,12530.9%12.7%
11Claude Haiku 4.5記載なし07/9015.7%73510.2%0.0%
公式・3 モデル・3 件 × 4 グループ(カバレッジ 4/90 ベンチマーク)Humanity's Last Exam(ツールあり)、Terminal-Bench-Science 0.1、OSWorld 2.1(partial)、Chartography(ツールあり)
  1. Humanity's Last Exam(ツールあり)公式・Introducing Claude Opus 5.5計測・報告: Anthropic(3)
  2. Terminal-Bench-Science 0.1公式・Introducing Claude Opus 5.5計測・報告: Anthropic(3)
  3. OSWorld 2.1(partial)公式・Introducing Claude Opus 5.5計測・報告: Anthropic(3)
  4. Chartography(ツールあり)公式・Introducing Claude Opus 5.5計測・報告: Anthropic(3)
順位モデルeffort平均百分位値のあるベンチマーク1. Humanity's Last Exam(ツールあり)2. Terminal-Bench-Science 0.13. OSWorld 2.1(partial)4. Chartography(ツールあり)
1Claude Opus 5.5Max10016/9067.7%58.7%81.8%89.0%
2Claude Fable 5.1記載なし5011/9065.6%52.6%80.7%88.4%
3Claude Opus 5記載なし04/9063.6%29.0%74.0%83.4%
公式・3 モデル・3 件 × 4 グループ(カバレッジ 4/90 ベンチマーク)GDPval-AA v2.1、Humanity's Last Exam(ツールあり)、OSWorld 2.1(partial)、Chartography(ツールなし)
  1. GDPval-AA v2.1公式・Introducing Claude Sonnet 5.5計測・報告: Artificial Analysis(3)
  2. Humanity's Last Exam(ツールあり)公式・Introducing Claude Sonnet 5.5計測・報告: Anthropic(3)
  3. OSWorld 2.1(partial)公式・Introducing Claude Sonnet 5.5計測・報告: Anthropic(3)
  4. Chartography(ツールなし)公式・Introducing Claude Sonnet 5.5計測・報告: Anthropic(3)
順位モデルeffort平均百分位値のあるベンチマーク1. GDPval-AA v2.12. Humanity's Last Exam(ツールあり)3. OSWorld 2.1(partial)4. Chartography(ツールなし)
1Claude Opus 5.5記載なし10014/901,84667.7%81.8%64.4%
2Claude Sonnet 5.5記載なし505/901,84464.5%80.1%61.6%
3Claude Sonnet 5記載なし04/901,44954.9%57.0%15.6%
公式・3 モデル・3 件 × 3 グループ(カバレッジ 3/90 ベンチマーク)AA-Briefcase v1.1(Elo)、Humanity's Last Exam(ツールあり)、Chartography(ツールなし)
  1. AA-Briefcase v1.1(Elo)公式・Introducing Claude Haiku 5.5計測・報告: Anthropic(3)
  2. Humanity's Last Exam(ツールあり)公式・Introducing Claude Haiku 5.5計測・報告: Anthropic(3)
  3. Chartography(ツールなし)公式・Introducing Claude Haiku 5.5計測・報告: Anthropic(3)
順位モデルeffort平均百分位値のあるベンチマーク1. AA-Briefcase v1.1(Elo)2. Humanity's Last Exam(ツールあり)3. Chartography(ツールなし)
1Claude Sonnet 5.5記載なし1005/901,82464.5%61.6%
2Claude Haiku 5.5記載なし504/901,57857.4%46.4%
3Claude Haiku 4.5記載なし07/9061418.7%6.4%

第三者

第三者・37 モデル・58 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)GPQA Diamond、OTIS Mock AIME 2024-2025、Chess Puzzles(Epoch AI)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(58)
  2. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(58)
  3. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(58)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. OTIS Mock AIME 2024-20253. Chess Puzzles(Epoch AI)
1Gemini 4 CarbonHigh978/1496.5%100.0%78.0%
2GPT-6 AstraMax959/1495.8%100.0%74.0%
3GPT-6.1 SolMax939/1495.4%100.0%68.0%
4Gemini 4 ArgonHigh939/1496.4%100.0%59.0%
5Claude Sonnet 5.5Max927/1495.6%100.0%59.0%
6GPT-6 SolMax877/1494.3%100.0%56.0%
7Claude Fable 5.2Max876/1493.9%99.7%79.0%
8GPT-5.6 SolMax866/1493.5%100.0%59.0%
9Gemini 3.8 FlashHigh827/1495.4%98.9%58.0%
10Kimi K3.1 ProMax816/1494.2%100.0%45.0%
11Claude Opus 5.5Max807/1490.6%100.0%67.0%
12GPT-5.6 TerraMax777/1493.3%99.7%47.0%
12Gemini 3.7 FlashHigh777/1494.8%97.2%66.0%
14DeepSeek-V4.1-Flash-1017Max736/1493.2%99.4%46.0%
15Qwen3.8-MaxXhigh718/1492.7%99.4%46.0%
16Gemini 3.1 Pro PreviewHigh716/1494.4%95.6%51.0%
17Claude Opus 4.8Max688/1491.0%98.3%52.0%
18DeepSeek-V4-Pro-0813Max677/1491.7%98.6%46.0%
19Grok 4.6Xhigh637/1493.2%99.2%33.0%
19GPT-5.6 LunaMax637/1491.6%98.3%43.0%
21Grok 4.7Xhigh627/1492.7%98.1%41.0%
21DeepSeek-V4.1-Flash記載なし628/1489.8%98.3%48.0%
23Claude Fable 5Max627/1485.9%99.7%47.0%
24Claude Fable 5High604/1483.3%100.0%45.0%
24Gemini 3.6 FlashHigh607/1494.1%94.2%40.0%
26Grok 4.6High594/1494.0%97.8%33.0%
27GPT-6 LunaMax579/1490.5%98.9%36.0%
28Kimi K3Max568/1493.1%97.2%36.0%
29Claude Opus 4.7Xhigh524/1490.2%97.8%36.0%
29Kimi K3High523/1491.9%93.3%37.0%
31GLM-5.3Max496/1490.9%91.1%37.0%
32Mistral Large 4High487/1488.1%98.3%33.0%
32Gemini 3 Flash PreviewHigh488/1489.4%95.6%40.0%
34GPT-5.6 SolLow443/1489.9%95.6%31.0%
35Claude Opus 4.8Low433/1488.4%97.8%29.0%
36Claude Opus 4.7Max426/1486.4%86.7%41.0%
37Gemini 3.6 FlashLow403/1486.4%82.2%42.0%
38GPT-5.5Low403/1490.7%84.4%30.0%
39GLM-5.3-FlashMax385/1490.2%93.9%21.0%
40Claude Fable 5Low353/1478.8%97.8%31.0%
41Claude Opus 4.6Max348/1488.4%91.1%21.0%
42GPT-5.6 TerraLow333/1487.4%88.9%27.0%
43Gemini 3.6 FlashMinimal293/1485.9%80.0%29.0%
44Claude Opus 4.8記載なし264/1485.4%84.4%19.0%
45Claude Sonnet 4.6High214/1483.3%75.6%19.0%
46Claude Sonnet 4.6Med203/1483.3%82.2%16.0%
47Kimi K3Low203/1484.9%68.9%18.0%
48Claude Sonnet 4.6Max194/1478.8%71.1%23.0%
49Gemini 3.5 Flash-LiteHigh185/1483.3%71.1%16.0%
50GPT-5.6 LunaLow154/1482.3%66.7%17.0%
51GPT-5.6 SolNone133/1482.8%68.9%7.0%
51Gemini 3.5 Flash-LiteLow133/1475.8%60.0%21.0%
53GPT-5.5None93/1477.3%57.8%9.0%
54Claude Opus 4.5記載なし74/1480.7%48.1%5.0%
55GPT-5.6 TerraNone63/1477.3%53.3%5.0%
56Gemini 3.5 Flash-LiteMinimal63/1474.2%51.1%7.0%
57Claude Sonnet 4.5記載なし29/1473.7%35.6%6.0%
58GPT-5.6 LunaNone14/1463.6%40.0%3.0%
第三者・36 モデル・40 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、OTIS Mock AIME 2024-2025
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(40)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(40)
  3. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(40)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. OTIS Mock AIME 2024-2025
1GPT-6 AstraMax939/1495.8%93.7%100.0%
2Gemini 4 CarbonHigh928/1496.5%90.5%100.0%
3GPT-6.1 SolMax919/1495.4%93.7%100.0%
4Claude Sonnet 5.5Max877/1495.6%88.8%100.0%
5GPT-6.1 TerraMax866/1494.3%90.9%100.0%
5Gemini 4 ArgonHigh869/1496.4%86.0%100.0%
7GPT-6 SolMax857/1494.3%89.8%100.0%
8Claude Fable 5.2Max806/1493.9%95.1%99.7%
9GPT-5.6 SolMax796/1493.5%89.1%100.0%
10Kimi K3.1 ProMax766/1494.2%78.3%100.0%
11Claude Opus 5.5Max737/1490.6%91.2%100.0%
12GPT-5.6 TerraMax707/1493.3%86.0%99.7%
13Gemini 3.8 FlashHigh637/1495.4%68.4%98.9%
14DeepSeek-V4.1-Flash-1017Max596/1493.2%71.2%99.4%
15Gemini 3.7 FlashHigh597/1494.8%71.6%97.2%
15Qwen3.8-MaxXhigh598/1492.7%74.7%99.4%
17Claude Fable 5Max567/1485.9%87.0%99.7%
18GPT-5.6 LunaMax547/1491.6%82.1%98.3%
19Grok 4.6Xhigh537/1493.2%66.0%99.2%
20GPT-6 LunaMax539/1490.5%79.0%98.9%
20Claude Opus 4.8Max538/1491.0%80.0%98.3%
22Kimi K3Max508/1493.1%72.2%97.2%
23Gemini 3.1 Pro Preview記載なし455/1494.1%59.7%95.6%
23DeepSeek-V4-Pro-0813Max457/1491.7%64.6%98.6%
25Gemini 3.6 FlashHigh447/1494.1%59.0%94.2%
26DeepSeek-V4.1-Flash記載なし408/1489.8%67.4%98.3%
27Grok 4.7Xhigh387/1492.7%53.0%98.1%
28GLM-5.3Max366/1490.9%68.8%91.1%
29Mistral Large 4High337/1488.1%58.6%98.3%
30Claude Opus 4.7Max306/1486.4%70.2%86.7%
31Claude Opus 4.6Max298/1488.4%66.0%91.1%
32GLM-5.3-FlashMax285/1490.2%55.8%93.9%
33Gemini 3 Flash Preview記載なし195/1483.2%51.2%92.8%
34Gemini 2.5 Pro記載なし135/1485.3%24.6%84.7%
35GPT-5.6 LunaLow114/1482.3%41.4%66.7%
35Gemini 3.5 Flash-LiteHigh115/1483.3%26.0%71.1%
37Gemini 3.1 Flash-LiteHigh103/1481.8%27.7%80.0%
38GPT-5.6 LunaNone54/1463.6%39.7%40.0%
39Gemini 3.1 Flash-LiteLow43/1474.2%22.5%44.4%
40Gemini 3.1 Flash-LiteMinimal13/1473.7%21.4%37.8%
第三者・35 モデル・35 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、OTIS Mock AIME 2024-2025
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(35)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(35)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(35)
  4. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(35)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. OTIS Mock AIME 2024-2025
1GPT-6 AstraMax939/1495.8%93.7%97.6%100.0%
2GPT-6.1 SolMax929/1495.4%93.7%100.0%100.0%
3Gemini 4 CarbonHigh918/1496.5%90.5%95.1%100.0%
4GPT-6.1 TerraMax856/1494.3%90.9%92.7%100.0%
5Claude Fable 5.2Max836/1493.9%95.1%100.0%99.7%
6Claude Sonnet 5.5Max827/1495.6%88.8%80.5%100.0%
7GPT-6 SolMax827/1494.3%89.8%90.0%100.0%
8Gemini 4 ArgonHigh819/1496.4%86.0%78.1%100.0%
9GPT-5.6 SolMax766/1493.5%89.1%82.9%100.0%
10Claude Opus 5.5Max747/1490.6%91.2%95.0%100.0%
11Kimi K3.1 ProMax706/1494.2%78.3%65.9%100.0%
12GPT-5.6 TerraMax667/1493.3%86.0%70.7%99.7%
13Claude Fable 5Max587/1485.9%87.0%90.2%99.7%
14DeepSeek-V4.1-Flash-1017Max536/1493.2%71.2%46.3%99.4%
15Qwen3.8-MaxXhigh528/1492.7%74.7%46.3%99.4%
16GPT-5.6 LunaMax517/1491.6%82.1%61.0%98.3%
17Gemini 3.7 FlashHigh507/1494.8%71.6%36.6%97.2%
18GPT-6 LunaMax499/1490.5%79.0%56.1%98.9%
18Claude Opus 4.8Max498/1491.0%80.0%56.1%98.3%
20Gemini 3.8 FlashHigh477/1495.4%68.4%22.0%98.9%
21Grok 4.6Xhigh447/1493.2%66.0%31.7%99.2%
22Kimi K3Max448/1493.1%72.2%39.0%97.2%
23Gemini 3.1 Pro Preview記載なし355/1494.1%59.7%26.8%95.6%
23DeepSeek-V4-Pro-0813Max357/1491.7%64.6%26.8%98.6%
25Gemini 3.6 FlashHigh317/1494.1%59.0%22.0%94.2%
26DeepSeek-V4.1-Flash記載なし308/1489.8%67.4%26.8%98.3%
27GLM-5.3Max296/1490.9%68.8%29.3%91.1%
28Claude Opus 4.7Max256/1486.4%70.2%31.7%86.7%
29Grok 4.7Xhigh247/1492.7%53.0%17.1%98.1%
30Mistral Large 4High247/1488.1%58.6%26.8%98.3%
31Claude Opus 4.6Max218/1488.4%66.0%26.8%91.1%
32GLM-5.3-FlashMax155/1490.2%55.8%17.1%93.9%
33Gemini 3 Flash Preview記載なし75/1483.2%51.2%17.1%92.8%
34Gemini 2.5 Pro記載なし35/1485.3%24.6%0.0%84.7%
35Gemini 3.5 Flash-LiteHigh25/1483.3%26.0%0.0%71.1%
第三者・33 モデル・35 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)GPQA Diamond、OTIS Mock AIME 2024-2025、SimpleQA Verified
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(35)
  2. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(35)
  3. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(35)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. OTIS Mock AIME 2024-20253. SimpleQA Verified
1Gemini 4 CarbonHigh958/1496.5%100.0%77.7%
2Gemini 4 ArgonHigh939/1496.4%100.0%76.0%
3GPT-6 AstraMax919/1495.8%100.0%75.6%
4GPT-6.1 SolMax879/1495.4%100.0%73.9%
5Claude Fable 5.2Max776/1493.9%99.7%94.0%
6GPT-6 SolMax757/1494.3%100.0%60.7%
6GPT-5.6 SolMax756/1493.5%100.0%69.7%
8Kimi K3.1 ProMax746/1494.2%100.0%62.2%
9GPT-6.1 TerraMax746/1494.3%100.0%57.3%
10Gemini 3.8 FlashHigh737/1495.4%98.9%69.7%
11Claude Sonnet 5.5Max707/1495.6%100.0%46.5%
12Claude Opus 5.5Max657/1490.6%100.0%72.2%
13Gemini 3.1 Pro PreviewHigh616/1494.4%95.6%73.5%
14Gemini 3.7 FlashHigh607/1494.8%97.2%69.2%
15GPT-5.6 TerraMax507/1493.3%99.7%43.2%
16Gemini 3.6 FlashHigh507/1494.1%94.2%66.2%
17Grok 4.6Xhigh507/1493.2%99.2%48.9%
18DeepSeek-V4-Pro-0813Max467/1491.7%98.6%52.9%
19Grok 4.7Xhigh457/1492.7%98.1%56.0%
19Grok 4.6High454/1494.0%97.8%49.3%
19DeepSeek-V4.1-Flash-1017Max456/1493.2%99.4%42.5%
22Qwen3.8-MaxXhigh458/1492.7%99.4%45.8%
23Claude Opus 4.8Max438/1491.0%98.3%53.0%
24Kimi K3Max398/1493.1%97.2%50.6%
25Gemini 3 Flash PreviewHigh358/1489.4%95.6%66.8%
26Claude Opus 4.7Xhigh334/1490.2%97.8%51.7%
27GPT-6 LunaMax329/1490.5%98.9%41.4%
28GPT-5.6 LunaMax307/1491.6%98.3%41.0%
29DeepSeek-V4.1-Flash記載なし288/1489.8%98.3%42.9%
30Claude Opus 4.6Max198/1488.4%91.1%47.0%
31GLM-5.3Max186/1490.9%91.1%41.0%
32Mistral Large 4High177/1488.1%98.3%20.0%
33Claude Sonnet 4.6High74/1483.3%75.6%35.5%
34Claude Sonnet 4.6Max44/1478.8%71.1%32.8%
35Claude Sonnet 4.5記載なし19/1473.7%35.6%23.7%
第三者・33 モデル・35 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)GPQA Diamond、SimpleQA Verified、Chess Puzzles(Epoch AI)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(35)
  2. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(35)
  3. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(35)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. SimpleQA Verified3. Chess Puzzles(Epoch AI)
1Gemini 4 CarbonHigh988/1496.5%77.7%78.0%
2GPT-6 AstraMax939/1495.8%75.6%74.0%
3Gemini 4 ArgonHigh909/1496.4%76.0%59.0%
4GPT-6.1 SolMax899/1495.4%73.9%68.0%
5Claude Fable 5.2Max886/1493.9%94.0%79.0%
6Gemini 3.7 FlashHigh807/1494.8%69.2%66.0%
7Gemini 3.8 FlashHigh797/1495.4%69.7%58.0%
8Gemini 3.1 Pro PreviewHigh756/1494.4%73.5%51.0%
9GPT-5.6 SolMax736/1493.5%69.7%59.0%
10GPT-6 SolMax707/1494.3%60.7%56.0%
11Claude Sonnet 5.5Max697/1495.6%46.5%59.0%
12Claude Opus 5.5Max677/1490.6%72.2%67.0%
13Kimi K3.1 ProMax616/1494.2%62.2%45.0%
14Gemini 3.6 FlashHigh577/1494.1%66.2%40.0%
15Claude Opus 4.8Max528/1491.0%53.0%52.0%
16Grok 4.7Xhigh487/1492.7%56.0%41.0%
16GPT-5.6 TerraMax487/1493.3%43.2%47.0%
16DeepSeek-V4-Pro-0813Max487/1491.7%52.9%46.0%
19DeepSeek-V4.1-Flash-1017Max436/1493.2%42.5%46.0%
20Grok 4.6High424/1494.0%49.3%33.0%
20Qwen3.8-MaxXhigh428/1492.7%45.8%46.0%
22Kimi K3Max408/1493.1%50.6%36.0%
23Gemini 3 Flash PreviewHigh408/1489.4%66.8%40.0%
24Grok 4.6Xhigh377/1493.2%48.9%33.0%
25DeepSeek-V4.1-Flash記載なし358/1489.8%42.9%48.0%
26Claude Opus 4.7Xhigh324/1490.2%51.7%36.0%
27GPT-5.6 LunaMax327/1491.6%41.0%43.0%
28Claude Haiku 5.5Max306/1489.7%23.8%55.0%
29GLM-5.3Max266/1490.9%41.0%37.0%
30GPT-6 LunaMax249/1490.5%41.4%36.0%
31Claude Opus 4.6Max198/1488.4%47.0%21.0%
32Mistral Large 4High87/1488.1%20.0%33.0%
33Claude Sonnet 4.6Max74/1478.8%32.8%23.0%
33Claude Sonnet 4.6High74/1483.3%35.5%19.0%
35Claude Sonnet 4.5記載なし19/1473.7%23.7%6.0%
第三者・33 モデル・33 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、Chess Puzzles(Epoch AI)
  1. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(33)
  2. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(33)
  3. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(33)
順位モデルeffort平均百分位値のあるベンチマーク1. FrontierMath Tiers 1-3(v2)2. FrontierMath Tier 4(v2)3. Chess Puzzles(Epoch AI)
1Claude Fable 5.2Max996/1495.1%100.0%79.0%
2GPT-6.1 SolMax959/1493.7%100.0%68.0%
3GPT-6 AstraMax949/1493.7%97.6%74.0%
4Gemini 4 CarbonHigh928/1490.5%95.1%78.0%
5Claude Opus 5.5Max897/1491.2%95.0%67.0%
6GPT-5.6 SolMax786/1489.1%82.9%59.0%
7GPT-6 SolMax777/1489.8%90.0%56.0%
8Claude Sonnet 5.5Max767/1488.8%80.5%59.0%
9Gemini 4 ArgonHigh729/1486.0%78.1%59.0%
10GPT-5.5Xhigh726/1485.3%72.5%63.0%
11Claude Fable 5Max707/1487.0%90.2%47.0%
12GPT-5.6 TerraMax637/1486.0%70.7%47.0%
13Claude Opus 4.8Max588/1480.0%56.1%52.0%
14Gemini 3.7 FlashHigh547/1471.6%36.6%66.0%
15Claude Haiku 5.5Max536/1475.1%46.3%55.0%
16GPT-5.6 LunaMax527/1482.1%61.0%43.0%
17Kimi K3.1 ProMax516/1478.3%65.9%45.0%
18Qwen3.8-MaxXhigh468/1474.7%46.3%46.0%
19GPT-6 LunaMax439/1479.0%56.1%36.0%
19DeepSeek-V4.1-Flash-1017Max436/1471.2%46.3%46.0%
21Gemini 3.8 FlashHigh367/1468.4%22.0%58.0%
22Kimi K3Max348/1472.2%39.0%36.0%
22DeepSeek-V4.1-Flash記載なし348/1467.4%26.8%48.0%
24Claude Opus 4.7Max326/1470.2%31.7%41.0%
25GLM-5.3Max276/1468.8%29.3%37.0%
26DeepSeek-V4-Pro-0813Max277/1464.6%26.8%46.0%
27Grok 4.6Xhigh217/1466.0%31.7%33.0%
28Gemini 3.6 FlashHigh167/1459.0%22.0%40.0%
29Claude Opus 4.6Max158/1466.0%26.8%21.0%
30Mistral Large 4High147/1458.6%26.8%33.0%
31Grok 4.7Xhigh137/1453.0%17.1%41.0%
32GLM-5.3-FlashMax55/1455.8%17.1%21.0%
33Gemini 3.5 Flash-LiteHigh05/1426.0%0.0%16.0%
第三者・32 モデル・34 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)GPQA Diamond、OTIS Mock AIME 2024-2025、SimpleQA Verified、Chess Puzzles(Epoch AI)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(34)
  2. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(34)
  3. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(34)
  4. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(34)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. OTIS Mock AIME 2024-20253. SimpleQA Verified4. Chess Puzzles(Epoch AI)
1Gemini 4 CarbonHigh958/1496.5%100.0%77.7%78.0%
2GPT-6 AstraMax929/1495.8%100.0%75.6%74.0%
3Gemini 4 ArgonHigh899/1496.4%100.0%76.0%59.0%
4GPT-6.1 SolMax889/1495.4%100.0%73.9%68.0%
5Claude Fable 5.2Max846/1493.9%99.7%94.0%79.0%
6GPT-5.6 SolMax766/1493.5%100.0%69.7%59.0%
7GPT-6 SolMax737/1494.3%100.0%60.7%56.0%
8Gemini 3.8 FlashHigh737/1495.4%98.9%69.7%58.0%
9Claude Sonnet 5.5Max737/1495.6%100.0%46.5%59.0%
10Claude Opus 5.5Max717/1490.6%100.0%72.2%67.0%
11Kimi K3.1 ProMax676/1494.2%100.0%62.2%45.0%
12Gemini 3.7 FlashHigh667/1494.8%97.2%69.2%66.0%
13Gemini 3.1 Pro PreviewHigh626/1494.4%95.6%73.5%51.0%
14GPT-5.6 TerraMax537/1493.3%99.7%43.2%47.0%
15Claude Opus 4.8Max508/1491.0%98.3%53.0%52.0%
16DeepSeek-V4-Pro-0813Max487/1491.7%98.6%52.9%46.0%
17DeepSeek-V4.1-Flash-1017Max486/1493.2%99.4%42.5%46.0%
18Qwen3.8-MaxXhigh478/1492.7%99.4%45.8%46.0%
19Gemini 3.6 FlashHigh477/1494.1%94.2%66.2%40.0%
20Grok 4.7Xhigh457/1492.7%98.1%56.0%41.0%
21Grok 4.6Xhigh427/1493.2%99.2%48.9%33.0%
22Grok 4.6High394/1494.0%97.8%49.3%33.0%
23DeepSeek-V4.1-Flash記載なし378/1489.8%98.3%42.9%48.0%
24Kimi K3Max368/1493.1%97.2%50.6%36.0%
25Gemini 3 Flash PreviewHigh358/1489.4%95.6%66.8%40.0%
26GPT-5.6 LunaMax347/1491.6%98.3%41.0%43.0%
27Claude Opus 4.7Xhigh324/1490.2%97.8%51.7%36.0%
28GPT-6 LunaMax319/1490.5%98.9%41.4%36.0%
29GLM-5.3Max216/1490.9%91.1%41.0%37.0%
30Mistral Large 4High177/1488.1%98.3%20.0%33.0%
31Claude Opus 4.6Max168/1488.4%91.1%47.0%21.0%
32Claude Sonnet 4.6High64/1483.3%75.6%35.5%19.0%
33Claude Sonnet 4.6Max54/1478.8%71.1%32.8%23.0%
34Claude Sonnet 4.5記載なし19/1473.7%35.6%23.7%6.0%
第三者・32 モデル・32 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、Chess Puzzles(Epoch AI)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(32)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(32)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(32)
  4. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(32)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. Chess Puzzles(Epoch AI)
1GPT-6 AstraMax949/1495.8%93.7%97.6%74.0%
2Gemini 4 CarbonHigh948/1496.5%90.5%95.1%78.0%
3GPT-6.1 SolMax929/1495.4%93.7%100.0%68.0%
4Claude Fable 5.2Max926/1493.9%95.1%100.0%79.0%
5Claude Sonnet 5.5Max807/1495.6%88.8%80.5%59.0%
6Gemini 4 ArgonHigh799/1496.4%86.0%78.1%59.0%
7GPT-6 SolMax777/1494.3%89.8%90.0%56.0%
8GPT-5.6 SolMax756/1493.5%89.1%82.9%59.0%
9Claude Opus 5.5Max737/1490.6%91.2%95.0%67.0%
10GPT-5.6 TerraMax637/1493.3%86.0%70.7%47.0%
11Gemini 3.7 FlashHigh617/1494.8%71.6%36.6%66.0%
12Kimi K3.1 ProMax586/1494.2%78.3%65.9%45.0%
13Claude Fable 5Max547/1485.9%87.0%90.2%47.0%
13Claude Opus 4.8Max548/1491.0%80.0%56.1%52.0%
15GPT-5.6 LunaMax507/1491.6%82.1%61.0%43.0%
16Gemini 3.8 FlashHigh497/1495.4%68.4%22.0%58.0%
17DeepSeek-V4.1-Flash-1017Max476/1493.2%71.2%46.3%46.0%
18Qwen3.8-MaxXhigh478/1492.7%74.7%46.3%46.0%
19Claude Haiku 5.5Max456/1489.7%75.1%46.3%55.0%
20GPT-6 LunaMax409/1490.5%79.0%56.1%36.0%
21Kimi K3Max398/1493.1%72.2%39.0%36.0%
22DeepSeek-V4-Pro-0813Max317/1491.7%64.6%26.8%46.0%
22DeepSeek-V4.1-Flash記載なし318/1489.8%67.4%26.8%48.0%
24Grok 4.6Xhigh317/1493.2%66.0%31.7%33.0%
25Gemini 3.6 FlashHigh307/1494.1%59.0%22.0%40.0%
26GLM-5.3Max296/1490.9%68.8%29.3%37.0%
27Claude Opus 4.7Max276/1486.4%70.2%31.7%41.0%
28Grok 4.7Xhigh227/1492.7%53.0%17.1%41.0%
29Claude Opus 4.6Max158/1488.4%66.0%26.8%21.0%
30Mistral Large 4High137/1488.1%58.6%26.8%33.0%
31GLM-5.3-FlashMax105/1490.2%55.8%17.1%21.0%
32Gemini 3.5 Flash-LiteHigh05/1483.3%26.0%0.0%16.0%
第三者・32 モデル・34 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、Chess Puzzles(Epoch AI)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(34)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(34)
  3. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(34)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. Chess Puzzles(Epoch AI)
1Gemini 4 CarbonHigh958/1496.5%90.5%78.0%
2GPT-6 AstraMax949/1495.8%93.7%74.0%
3GPT-6.1 SolMax919/1495.4%93.7%68.0%
4Claude Fable 5.2Max906/1493.9%95.1%79.0%
5Claude Sonnet 5.5Max837/1495.6%88.8%59.0%
6Gemini 4 ArgonHigh829/1496.4%86.0%59.0%
7GPT-6 SolMax787/1494.3%89.8%56.0%
8GPT-5.6 SolMax766/1493.5%89.1%59.0%
9Claude Opus 5.5Max717/1490.6%91.2%67.0%
9Gemini 3.7 FlashHigh717/1494.8%71.6%66.0%
11Gemini 3.8 FlashHigh647/1495.4%68.4%58.0%
12GPT-5.6 TerraMax647/1493.3%86.0%47.0%
13Kimi K3.1 ProMax596/1494.2%78.3%45.0%
14Claude Opus 4.8Max568/1491.0%80.0%52.0%
15DeepSeek-V4.1-Flash-1017Max506/1493.2%71.2%46.0%
16GPT-5.6 LunaMax497/1491.6%82.1%43.0%
16Qwen3.8-MaxXhigh498/1492.7%74.7%46.0%
18Claude Haiku 5.5Max476/1489.7%75.1%55.0%
19Claude Fable 5Max477/1485.9%87.0%47.0%
20Kimi K3Max428/1493.1%72.2%36.0%
21Gemini 3.6 FlashHigh407/1494.1%59.0%40.0%
22DeepSeek-V4-Pro-0813Max387/1491.7%64.6%46.0%
22DeepSeek-V4.1-Flash記載なし388/1489.8%67.4%48.0%
24GPT-6 LunaMax389/1490.5%79.0%36.0%
25Grok 4.6Xhigh347/1493.2%66.0%33.0%
26GLM-5.3Max336/1490.9%68.8%37.0%
27Grok 4.7Xhigh327/1492.7%53.0%41.0%
28Claude Opus 4.7Max296/1486.4%70.2%41.0%
29Claude Opus 4.6Max188/1488.4%66.0%21.0%
30GLM-5.3-FlashMax175/1490.2%55.8%21.0%
31Mistral Large 4High167/1488.1%58.6%33.0%
32GPT-5.6 LunaLow54/1482.3%41.4%17.0%
33Gemini 3.5 Flash-LiteHigh35/1483.3%26.0%16.0%
34GPT-5.6 LunaNone14/1463.6%39.7%3.0%
第三者・31 モデル・31 件 × 5 グループ(カバレッジ 5/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、OTIS Mock AIME 2024-2025、Chess Puzzles(Epoch AI)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(31)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(31)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(31)
  4. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(31)
  5. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(31)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. OTIS Mock AIME 2024-20255. Chess Puzzles(Epoch AI)
1GPT-6 AstraMax929/1495.8%93.7%97.6%100.0%74.0%
2Gemini 4 CarbonHigh928/1496.5%90.5%95.1%100.0%78.0%
3GPT-6.1 SolMax919/1495.4%93.7%100.0%100.0%68.0%
4Claude Fable 5.2Max866/1493.9%95.1%100.0%99.7%79.0%
5Claude Sonnet 5.5Max817/1495.6%88.8%80.5%100.0%59.0%
6Gemini 4 ArgonHigh809/1496.4%86.0%78.1%100.0%59.0%
7GPT-6 SolMax797/1494.3%89.8%90.0%100.0%56.0%
8GPT-5.6 SolMax776/1493.5%89.1%82.9%100.0%59.0%
9Claude Opus 5.5Max757/1490.6%91.2%95.0%100.0%67.0%
10Kimi K3.1 ProMax636/1494.2%78.3%65.9%100.0%45.0%
10GPT-5.6 TerraMax637/1493.3%86.0%70.7%99.7%47.0%
12Claude Fable 5Max567/1485.9%87.0%90.2%99.7%47.0%
13Gemini 3.7 FlashHigh547/1494.8%71.6%36.6%97.2%66.0%
14DeepSeek-V4.1-Flash-1017Max506/1493.2%71.2%46.3%99.4%46.0%
15Claude Opus 4.8Max498/1491.0%80.0%56.1%98.3%52.0%
15Qwen3.8-MaxXhigh498/1492.7%74.7%46.3%99.4%46.0%
17Gemini 3.8 FlashHigh497/1495.4%68.4%22.0%98.9%58.0%
18GPT-5.6 LunaMax467/1491.6%82.1%61.0%98.3%43.0%
19GPT-6 LunaMax409/1490.5%79.0%56.1%98.9%36.0%
20Kimi K3Max368/1493.1%72.2%39.0%97.2%36.0%
21Grok 4.6Xhigh357/1493.2%66.0%31.7%99.2%33.0%
22DeepSeek-V4-Pro-0813Max347/1491.7%64.6%26.8%98.6%46.0%
23DeepSeek-V4.1-Flash記載なし328/1489.8%67.4%26.8%98.3%48.0%
24Gemini 3.6 FlashHigh287/1494.1%59.0%22.0%94.2%40.0%
25GLM-5.3Max256/1490.9%68.8%29.3%91.1%37.0%
26Grok 4.7Xhigh237/1492.7%53.0%17.1%98.1%41.0%
26Claude Opus 4.7Max236/1486.4%70.2%31.7%86.7%41.0%
28Mistral Large 4High187/1488.1%58.6%26.8%98.3%33.0%
29Claude Opus 4.6Max148/1488.4%66.0%26.8%91.1%21.0%
30GLM-5.3-FlashMax105/1490.2%55.8%17.1%93.9%21.0%
31Gemini 3.5 Flash-LiteHigh05/1483.3%26.0%0.0%71.1%16.0%
第三者・31 モデル・33 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、OTIS Mock AIME 2024-2025、Chess Puzzles(Epoch AI)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(33)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(33)
  3. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(33)
  4. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(33)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. OTIS Mock AIME 2024-20254. Chess Puzzles(Epoch AI)
1Gemini 4 CarbonHigh938/1496.5%90.5%100.0%78.0%
2GPT-6 AstraMax939/1495.8%93.7%100.0%74.0%
3GPT-6.1 SolMax909/1495.4%93.7%100.0%68.0%
4Claude Fable 5.2Max846/1493.9%95.1%99.7%79.0%
5Claude Sonnet 5.5Max847/1495.6%88.8%100.0%59.0%
6Gemini 4 ArgonHigh839/1496.4%86.0%100.0%59.0%
7GPT-6 SolMax807/1494.3%89.8%100.0%56.0%
8GPT-5.6 SolMax786/1493.5%89.1%100.0%59.0%
9Claude Opus 5.5Max747/1490.6%91.2%100.0%67.0%
10Kimi K3.1 ProMax666/1494.2%78.3%100.0%45.0%
11GPT-5.6 TerraMax657/1493.3%86.0%99.7%47.0%
12Gemini 3.8 FlashHigh617/1495.4%68.4%98.9%58.0%
13Gemini 3.7 FlashHigh607/1494.8%71.6%97.2%66.0%
14DeepSeek-V4.1-Flash-1017Max536/1493.2%71.2%99.4%46.0%
15Claude Fable 5Max537/1485.9%87.0%99.7%47.0%
15Qwen3.8-MaxXhigh538/1492.7%74.7%99.4%46.0%
17Claude Opus 4.8Max518/1491.0%80.0%98.3%52.0%
18GPT-5.6 LunaMax467/1491.6%82.1%98.3%43.0%
19GPT-6 LunaMax419/1490.5%79.0%98.9%36.0%
19DeepSeek-V4-Pro-0813Max417/1491.7%64.6%98.6%46.0%
21Grok 4.6Xhigh397/1493.2%66.0%99.2%33.0%
22DeepSeek-V4.1-Flash記載なし398/1489.8%67.4%98.3%48.0%
23Kimi K3Max388/1493.1%72.2%97.2%36.0%
24Gemini 3.6 FlashHigh367/1494.1%59.0%94.2%40.0%
25Grok 4.7Xhigh327/1492.7%53.0%98.1%41.0%
26GLM-5.3Max296/1490.9%68.8%91.1%37.0%
27Claude Opus 4.7Max256/1486.4%70.2%86.7%41.0%
28Mistral Large 4High227/1488.1%58.6%98.3%33.0%
29Claude Opus 4.6Max188/1488.4%66.0%91.1%21.0%
30GLM-5.3-FlashMax175/1490.2%55.8%93.9%21.0%
31GPT-5.6 LunaLow54/1482.3%41.4%66.7%17.0%
32Gemini 3.5 Flash-LiteHigh45/1483.3%26.0%71.1%16.0%
33GPT-5.6 LunaNone14/1463.6%39.7%40.0%3.0%
第三者・30 モデル・30 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、SimpleQA Verified
  1. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(30)
  2. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(30)
  3. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(30)
順位モデルeffort平均百分位値のあるベンチマーク1. FrontierMath Tiers 1-3(v2)2. FrontierMath Tier 4(v2)3. SimpleQA Verified
1Claude Fable 5.2Max996/1495.1%100.0%94.0%
2GPT-6.1 SolMax939/1493.7%100.0%73.9%
3GPT-6 AstraMax939/1493.7%97.6%75.6%
4Gemini 4 CarbonHigh908/1490.5%95.1%77.7%
5Claude Opus 5.5Max867/1491.2%95.0%72.2%
6Gemini 4 ArgonHigh769/1486.0%78.1%76.0%
6GPT-5.6 SolMax766/1489.1%82.9%69.7%
8GPT-6.1 TerraMax756/1490.9%92.7%57.3%
9GPT-6 SolMax727/1489.8%90.0%60.7%
10GPT-5.5Xhigh646/1485.3%72.5%63.0%
11Claude Sonnet 5.5Max597/1488.8%80.5%46.5%
12Kimi K3.1 ProMax566/1478.3%65.9%62.2%
13GPT-5.6 TerraMax517/1486.0%70.7%43.2%
13Claude Opus 4.8Max518/1480.0%56.1%53.0%
15Gemini 3.7 FlashHigh467/1471.6%36.6%69.2%
16GPT-5.6 LunaMax417/1482.1%61.0%41.0%
17GPT-6 LunaMax399/1479.0%56.1%41.4%
18Kimi K3Max388/1472.2%39.0%50.6%
19Qwen3.8-MaxXhigh378/1474.7%46.3%45.8%
20Gemini 3.8 FlashHigh367/1468.4%22.0%69.7%
21Claude Haiku 5.5Max306/1475.1%46.3%23.8%
21DeepSeek-V4.1-Flash-1017Max306/1471.2%46.3%42.5%
23Grok 4.6Xhigh277/1466.0%31.7%48.9%
23Gemini 3.6 FlashHigh277/1459.0%22.0%66.2%
25DeepSeek-V4-Pro-0813Max247/1464.6%26.8%52.9%
26Claude Opus 4.6Max228/1466.0%26.8%47.0%
27GLM-5.3Max206/1468.8%29.3%41.0%
28DeepSeek-V4.1-Flash記載なし198/1467.4%26.8%42.9%
29Grok 4.7Xhigh177/1453.0%17.1%56.0%
30Mistral Large 4High67/1458.6%26.8%20.0%
第三者・29 モデル・29 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、SimpleQA Verified
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(29)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(29)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(29)
  4. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(29)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. SimpleQA Verified
1GPT-6 AstraMax929/1495.8%93.7%97.6%75.6%
2Gemini 4 CarbonHigh928/1496.5%90.5%95.1%77.7%
3GPT-6.1 SolMax919/1495.4%93.7%100.0%73.9%
4Claude Fable 5.2Max906/1493.9%95.1%100.0%94.0%
5Gemini 4 ArgonHigh819/1496.4%86.0%78.1%76.0%
6GPT-6.1 TerraMax756/1494.3%90.9%92.7%57.3%
7GPT-6 SolMax737/1494.3%89.8%90.0%60.7%
8GPT-5.6 SolMax716/1493.5%89.1%82.9%69.7%
9Claude Opus 5.5Max697/1490.6%91.2%95.0%72.2%
10Claude Sonnet 5.5Max667/1495.6%88.8%80.5%46.5%
11Kimi K3.1 ProMax616/1494.2%78.3%65.9%62.2%
12Gemini 3.7 FlashHigh547/1494.8%71.6%36.6%69.2%
13GPT-5.6 TerraMax527/1493.3%86.0%70.7%43.2%
14Gemini 3.8 FlashHigh487/1495.4%68.4%22.0%69.7%
15Claude Opus 4.8Max468/1491.0%80.0%56.1%53.0%
16Kimi K3Max408/1493.1%72.2%39.0%50.6%
17GPT-5.6 LunaMax397/1491.6%82.1%61.0%41.0%
18Qwen3.8-MaxXhigh388/1492.7%74.7%46.3%45.8%
19Gemini 3.6 FlashHigh367/1494.1%59.0%22.0%66.2%
20DeepSeek-V4.1-Flash-1017Max356/1493.2%71.2%46.3%42.5%
21GPT-6 LunaMax339/1490.5%79.0%56.1%41.4%
22Grok 4.6Xhigh337/1493.2%66.0%31.7%48.9%
23DeepSeek-V4-Pro-0813Max267/1491.7%64.6%26.8%52.9%
24Claude Haiku 5.5Max256/1489.7%75.1%46.3%23.8%
25Grok 4.7Xhigh237/1492.7%53.0%17.1%56.0%
26GLM-5.3Max216/1490.9%68.8%29.3%41.0%
27Claude Opus 4.6Max188/1488.4%66.0%26.8%47.0%
28DeepSeek-V4.1-Flash記載なし178/1489.8%67.4%26.8%42.9%
29Mistral Large 4High57/1488.1%58.6%26.8%20.0%
第三者・29 モデル・29 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、SimpleQA Verified、Chess Puzzles(Epoch AI)
  1. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(29)
  2. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(29)
  3. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(29)
  4. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(29)
順位モデルeffort平均百分位値のあるベンチマーク1. FrontierMath Tiers 1-3(v2)2. FrontierMath Tier 4(v2)3. SimpleQA Verified4. Chess Puzzles(Epoch AI)
1Claude Fable 5.2Max1006/1495.1%100.0%94.0%79.0%
2GPT-6 AstraMax929/1493.7%97.6%75.6%74.0%
3Gemini 4 CarbonHigh928/1490.5%95.1%77.7%78.0%
3GPT-6.1 SolMax929/1493.7%100.0%73.9%68.0%
5Claude Opus 5.5Max867/1491.2%95.0%72.2%67.0%
6Gemini 4 ArgonHigh769/1486.0%78.1%76.0%59.0%
6GPT-5.6 SolMax766/1489.1%82.9%69.7%59.0%
8GPT-6 SolMax717/1489.8%90.0%60.7%56.0%
9GPT-5.5Xhigh696/1485.3%72.5%63.0%63.0%
10Claude Sonnet 5.5Max637/1488.8%80.5%46.5%59.0%
11Gemini 3.7 FlashHigh557/1471.6%36.6%69.2%66.0%
12Claude Opus 4.8Max538/1480.0%56.1%53.0%52.0%
13GPT-5.6 TerraMax517/1486.0%70.7%43.2%47.0%
14Kimi K3.1 ProMax516/1478.3%65.9%62.2%45.0%
15Gemini 3.8 FlashHigh437/1468.4%22.0%69.7%58.0%
16GPT-5.6 LunaMax397/1482.1%61.0%41.0%43.0%
17Qwen3.8-MaxXhigh388/1474.7%46.3%45.8%46.0%
18Claude Haiku 5.5Max386/1475.1%46.3%23.8%55.0%
19GPT-6 LunaMax339/1479.0%56.1%41.4%36.0%
19DeepSeek-V4.1-Flash-1017Max336/1471.2%46.3%42.5%46.0%
21Kimi K3Max338/1472.2%39.0%50.6%36.0%
22DeepSeek-V4-Pro-0813Max287/1464.6%26.8%52.9%46.0%
23DeepSeek-V4.1-Flash記載なし278/1467.4%26.8%42.9%48.0%
24Gemini 3.6 FlashHigh257/1459.0%22.0%66.2%40.0%
25Grok 4.6Xhigh227/1466.0%31.7%48.9%33.0%
26GLM-5.3Max206/1468.8%29.3%41.0%37.0%
27Grok 4.7Xhigh207/1453.0%17.1%56.0%41.0%
28Claude Opus 4.6Max178/1466.0%26.8%47.0%21.0%
29Mistral Large 4High67/1458.6%26.8%20.0%33.0%
第三者・28 モデル・28 件 × 5 グループ(カバレッジ 5/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、OTIS Mock AIME 2024-2025、SimpleQA Verified
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(28)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(28)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(28)
  4. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(28)
  5. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(28)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. OTIS Mock AIME 2024-20255. SimpleQA Verified
1GPT-6 AstraMax909/1495.8%93.7%97.6%100.0%75.6%
2Gemini 4 CarbonHigh908/1496.5%90.5%95.1%100.0%77.7%
3GPT-6.1 SolMax899/1495.4%93.7%100.0%100.0%73.9%
4Claude Fable 5.2Max846/1493.9%95.1%100.0%99.7%94.0%
5Gemini 4 ArgonHigh819/1496.4%86.0%78.1%100.0%76.0%
6GPT-6.1 TerraMax766/1494.3%90.9%92.7%100.0%57.3%
7GPT-6 SolMax747/1494.3%89.8%90.0%100.0%60.7%
8GPT-5.6 SolMax736/1493.5%89.1%82.9%100.0%69.7%
9Claude Opus 5.5Max717/1490.6%91.2%95.0%100.0%72.2%
10Claude Sonnet 5.5Max697/1495.6%88.8%80.5%100.0%46.5%
11Kimi K3.1 ProMax646/1494.2%78.3%65.9%100.0%62.2%
12GPT-5.6 TerraMax537/1493.3%86.0%70.7%99.7%43.2%
13Gemini 3.8 FlashHigh477/1495.4%68.4%22.0%98.9%69.7%
14Gemini 3.7 FlashHigh467/1494.8%71.6%36.6%97.2%69.2%
15Claude Opus 4.8Max418/1491.0%80.0%56.1%98.3%53.0%
16Qwen3.8-MaxXhigh408/1492.7%74.7%46.3%99.4%45.8%
17DeepSeek-V4.1-Flash-1017Max386/1493.2%71.2%46.3%99.4%42.5%
18Grok 4.6Xhigh367/1493.2%66.0%31.7%99.2%48.9%
19GPT-5.6 LunaMax357/1491.6%82.1%61.0%98.3%41.0%
20Kimi K3Max348/1493.1%72.2%39.0%97.2%50.6%
21GPT-6 LunaMax339/1490.5%79.0%56.1%98.9%41.4%
22Gemini 3.6 FlashHigh307/1494.1%59.0%22.0%94.2%66.2%
23DeepSeek-V4-Pro-0813Max287/1491.7%64.6%26.8%98.6%52.9%
24Grok 4.7Xhigh217/1492.7%53.0%17.1%98.1%56.0%
25DeepSeek-V4.1-Flash記載なし198/1489.8%67.4%26.8%98.3%42.9%
26GLM-5.3Max166/1490.9%68.8%29.3%91.1%41.0%
27Claude Opus 4.6Max148/1488.4%66.0%26.8%91.1%47.0%
28Mistral Large 4High107/1488.1%58.6%26.8%98.3%20.0%
第三者・28 モデル・28 件 × 5 グループ(カバレッジ 5/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、SimpleQA Verified、Chess Puzzles(Epoch AI)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(28)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(28)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(28)
  4. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(28)
  5. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(28)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. SimpleQA Verified5. Chess Puzzles(Epoch AI)
1Gemini 4 CarbonHigh938/1496.5%90.5%95.1%77.7%78.0%
2Claude Fable 5.2Max926/1493.9%95.1%100.0%94.0%79.0%
2GPT-6 AstraMax929/1495.8%93.7%97.6%75.6%74.0%
4GPT-6.1 SolMax909/1495.4%93.7%100.0%73.9%68.0%
5Gemini 4 ArgonHigh809/1496.4%86.0%78.1%76.0%59.0%
6GPT-5.6 SolMax736/1493.5%89.1%82.9%69.7%59.0%
7Claude Opus 5.5Max727/1490.6%91.2%95.0%72.2%67.0%
7GPT-6 SolMax727/1494.3%89.8%90.0%60.7%56.0%
9Claude Sonnet 5.5Max697/1495.6%88.8%80.5%46.5%59.0%
10Gemini 3.7 FlashHigh607/1494.8%71.6%36.6%69.2%66.0%
11Kimi K3.1 ProMax566/1494.2%78.3%65.9%62.2%45.0%
12GPT-5.6 TerraMax537/1493.3%86.0%70.7%43.2%47.0%
13Gemini 3.8 FlashHigh517/1495.4%68.4%22.0%69.7%58.0%
14Claude Opus 4.8Max498/1491.0%80.0%56.1%53.0%52.0%
15Qwen3.8-MaxXhigh398/1492.7%74.7%46.3%45.8%46.0%
16GPT-5.6 LunaMax387/1491.6%82.1%61.0%41.0%43.0%
17DeepSeek-V4.1-Flash-1017Max376/1493.2%71.2%46.3%42.5%46.0%
18Kimi K3Max368/1493.1%72.2%39.0%50.6%36.0%
19Gemini 3.6 FlashHigh347/1494.1%59.0%22.0%66.2%40.0%
20Claude Haiku 5.5Max336/1489.7%75.1%46.3%23.8%55.0%
21GPT-6 LunaMax309/1490.5%79.0%56.1%41.4%36.0%
22DeepSeek-V4-Pro-0813Max307/1491.7%64.6%26.8%52.9%46.0%
23Grok 4.6Xhigh297/1493.2%66.0%31.7%48.9%33.0%
24DeepSeek-V4.1-Flash記載なし258/1489.8%67.4%26.8%42.9%48.0%
25Grok 4.7Xhigh247/1492.7%53.0%17.1%56.0%41.0%
26GLM-5.3Max216/1490.9%68.8%29.3%41.0%37.0%
27Claude Opus 4.6Max158/1488.4%66.0%26.8%47.0%21.0%
28Mistral Large 4High57/1488.1%58.6%26.8%20.0%33.0%
第三者・27 モデル・27 件 × 6 グループ(カバレッジ 6/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、OTIS Mock AIME 2024-2025、SimpleQA Verified、Chess Puzzles(Epoch AI)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(27)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(27)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(27)
  4. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(27)
  5. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(27)
  6. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(27)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. OTIS Mock AIME 2024-20255. SimpleQA Verified6. Chess Puzzles(Epoch AI)
1Gemini 4 CarbonHigh928/1496.5%90.5%95.1%100.0%77.7%78.0%
2GPT-6 AstraMax919/1495.8%93.7%97.6%100.0%75.6%74.0%
3GPT-6.1 SolMax899/1495.4%93.7%100.0%100.0%73.9%68.0%
4Claude Fable 5.2Max876/1493.9%95.1%100.0%99.7%94.0%79.0%
5Gemini 4 ArgonHigh809/1496.4%86.0%78.1%100.0%76.0%59.0%
6GPT-5.6 SolMax746/1493.5%89.1%82.9%100.0%69.7%59.0%
7Claude Opus 5.5Max737/1490.6%91.2%95.0%100.0%72.2%67.0%
7GPT-6 SolMax737/1494.3%89.8%90.0%100.0%60.7%56.0%
9Claude Sonnet 5.5Max717/1495.6%88.8%80.5%100.0%46.5%59.0%
10Kimi K3.1 ProMax606/1494.2%78.3%65.9%100.0%62.2%45.0%
11GPT-5.6 TerraMax547/1493.3%86.0%70.7%99.7%43.2%47.0%
12Gemini 3.7 FlashHigh527/1494.8%71.6%36.6%97.2%69.2%66.0%
13Gemini 3.8 FlashHigh507/1495.4%68.4%22.0%98.9%69.7%58.0%
14Claude Opus 4.8Max458/1491.0%80.0%56.1%98.3%53.0%52.0%
15Qwen3.8-MaxXhigh428/1492.7%74.7%46.3%99.4%45.8%46.0%
16DeepSeek-V4.1-Flash-1017Max406/1493.2%71.2%46.3%99.4%42.5%46.0%
17GPT-5.6 LunaMax357/1491.6%82.1%61.0%98.3%41.0%43.0%
18Kimi K3Max328/1493.1%72.2%39.0%97.2%50.6%36.0%
19Grok 4.6Xhigh327/1493.2%66.0%31.7%99.2%48.9%33.0%
20GPT-6 LunaMax319/1490.5%79.0%56.1%98.9%41.4%36.0%
20DeepSeek-V4-Pro-0813Max317/1491.7%64.6%26.8%98.6%52.9%46.0%
22Gemini 3.6 FlashHigh297/1494.1%59.0%22.0%94.2%66.2%40.0%
23DeepSeek-V4.1-Flash記載なし258/1489.8%67.4%26.8%98.3%42.9%48.0%
24Grok 4.7Xhigh237/1492.7%53.0%17.1%98.1%56.0%41.0%
25GLM-5.3Max176/1490.9%68.8%29.3%91.1%41.0%37.0%
26Claude Opus 4.6Max138/1488.4%66.0%26.8%91.1%47.0%21.0%
27Mistral Large 4High97/1488.1%58.6%26.8%98.3%20.0%33.0%
第三者・25 モデル・25 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、MirrorCode
  1. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(25)
  2. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(25)
  3. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(25)
順位モデルeffort平均百分位値のあるベンチマーク1. FrontierMath Tiers 1-3(v2)2. FrontierMath Tier 4(v2)3. MirrorCode
1Claude Opus 5.5Max937/1491.2%95.0%95.6%
2GPT-6.1 SolMax909/1493.7%100.0%61.1%
3GPT-6.1 TerraMax886/1490.9%92.7%76.7%
4GPT-6 AstraMax879/1493.7%97.6%60.0%
5Claude Sonnet 5.5Max857/1488.8%80.5%97.8%
6Claude Fable 5Max817/1487.0%90.2%70.0%
7GPT-6 SolMax697/1489.8%90.0%48.9%
8Gemini 4 ArgonHigh679/1486.0%78.1%58.9%
9Claude Opus 4.8Max628/1480.0%56.1%65.6%
10Claude Haiku 5.5Max606/1475.1%46.3%78.9%
11GPT-5.6 TerraMax567/1486.0%70.7%41.1%
12Qwen3.8-MaxXhigh458/1474.7%46.3%52.2%
12GPT-5.5Xhigh456/1485.3%72.5%23.3%
14Kimi K3Max448/1472.2%39.0%54.4%
15DeepSeek-V4.1-Flash記載なし408/1467.4%26.8%64.4%
16GPT-5.6 LunaMax397/1482.1%61.0%21.1%
17GPT-6 LunaMax389/1479.0%56.1%26.7%
18Gemini 3.7 FlashHigh357/1471.6%36.6%41.1%
19Grok 4.6Xhigh287/1466.0%31.7%41.1%
20Gemini 3.8 FlashHigh227/1468.4%22.0%38.9%
21Claude Opus 4.6Max208/1466.0%26.8%33.3%
22Grok 4.7Xhigh197/1453.0%17.1%54.4%
23DeepSeek-V4-Pro-0813Max187/1464.6%26.8%33.3%
24Mistral Large 4High137/1458.6%26.8%27.8%
25Gemini 3.6 FlashHigh77/1459.0%22.0%23.3%
第三者・24 モデル・24 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(24)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(24)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(24)
  4. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(24)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. MirrorCode
1GPT-6 AstraMax899/1495.8%93.7%97.6%60.0%
2GPT-6.1 SolMax889/1495.4%93.7%100.0%61.1%
3Claude Sonnet 5.5Max867/1495.6%88.8%80.5%97.8%
4GPT-6.1 TerraMax836/1494.3%90.9%92.7%76.7%
5Claude Opus 5.5Max767/1490.6%91.2%95.0%95.6%
6Gemini 4 ArgonHigh749/1496.4%86.0%78.1%58.9%
7GPT-6 SolMax697/1494.3%89.8%90.0%48.9%
8Claude Fable 5Max607/1485.9%87.0%90.2%70.0%
9GPT-5.6 TerraMax577/1493.3%86.0%70.7%41.1%
10Claude Opus 4.8Max558/1491.0%80.0%56.1%65.6%
11Claude Haiku 5.5Max496/1489.7%75.1%46.3%78.9%
12Kimi K3Max468/1493.1%72.2%39.0%54.4%
13Gemini 3.7 FlashHigh467/1494.8%71.6%36.6%41.1%
14Qwen3.8-MaxXhigh458/1492.7%74.7%46.3%52.2%
15GPT-5.6 LunaMax397/1491.6%82.1%61.0%21.1%
16Gemini 3.8 FlashHigh377/1495.4%68.4%22.0%38.9%
17Grok 4.6Xhigh357/1493.2%66.0%31.7%41.1%
18GPT-6 LunaMax349/1490.5%79.0%56.1%26.7%
18DeepSeek-V4.1-Flash記載なし348/1489.8%67.4%26.8%64.4%
20Grok 4.7Xhigh267/1492.7%53.0%17.1%54.4%
21DeepSeek-V4-Pro-0813Max237/1491.7%64.6%26.8%33.3%
22Gemini 3.6 FlashHigh217/1494.1%59.0%22.0%23.3%
23Claude Opus 4.6Max178/1488.4%66.0%26.8%33.3%
24Mistral Large 4High107/1488.1%58.6%26.8%27.8%
第三者・24 モデル・24 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、SimpleQA Verified、MirrorCode
  1. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(24)
  2. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(24)
  3. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(24)
  4. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(24)
順位モデルeffort平均百分位値のあるベンチマーク1. FrontierMath Tiers 1-3(v2)2. FrontierMath Tier 4(v2)3. SimpleQA Verified4. MirrorCode
1Claude Opus 5.5Max917/1491.2%95.0%72.2%95.6%
2GPT-6.1 SolMax919/1493.7%100.0%73.9%61.1%
3GPT-6 AstraMax909/1493.7%97.6%75.6%60.0%
4GPT-6.1 TerraMax806/1490.9%92.7%57.3%76.7%
5Gemini 4 ArgonHigh789/1486.0%78.1%76.0%58.9%
6Claude Sonnet 5.5Max727/1488.8%80.5%46.5%97.8%
7GPT-6 SolMax707/1489.8%90.0%60.7%48.9%
8Claude Opus 4.8Max618/1480.0%56.1%53.0%65.6%
9GPT-5.5Xhigh536/1485.3%72.5%63.0%23.3%
10GPT-5.6 TerraMax497/1486.0%70.7%43.2%41.1%
11Claude Haiku 5.5Max476/1475.1%46.3%23.8%78.9%
12Gemini 3.7 FlashHigh477/1471.6%36.6%69.2%41.1%
13Kimi K3Max458/1472.2%39.0%50.6%54.4%
14Qwen3.8-MaxXhigh428/1474.7%46.3%45.8%52.2%
15Gemini 3.8 FlashHigh387/1468.4%22.0%69.7%38.9%
16DeepSeek-V4.1-Flash記載なし358/1467.4%26.8%42.9%64.4%
17GPT-6 LunaMax339/1479.0%56.1%41.4%26.7%
18GPT-5.6 LunaMax337/1482.1%61.0%41.0%21.1%
19Grok 4.6Xhigh327/1466.0%31.7%48.9%41.1%
20Grok 4.7Xhigh297/1453.0%17.1%56.0%54.4%
21DeepSeek-V4-Pro-0813Max267/1464.6%26.8%52.9%33.3%
22Claude Opus 4.6Max248/1466.0%26.8%47.0%33.3%
23Gemini 3.6 FlashHigh247/1459.0%22.0%66.2%23.3%
24Mistral Large 4High107/1458.6%26.8%20.0%27.8%
第三者・24 モデル・24 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、Chess Puzzles(Epoch AI)、MirrorCode
  1. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(24)
  2. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(24)
  3. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(24)
  4. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(24)
順位モデルeffort平均百分位値のあるベンチマーク1. FrontierMath Tiers 1-3(v2)2. FrontierMath Tier 4(v2)3. Chess Puzzles(Epoch AI)4. MirrorCode
1Claude Opus 5.5Max927/1491.2%95.0%67.0%95.6%
2GPT-6.1 SolMax929/1493.7%100.0%68.0%61.1%
3GPT-6 AstraMax919/1493.7%97.6%74.0%60.0%
4Claude Sonnet 5.5Max847/1488.8%80.5%59.0%97.8%
5Claude Fable 5Max747/1487.0%90.2%47.0%70.0%
6Gemini 4 ArgonHigh729/1486.0%78.1%59.0%58.9%
7GPT-6 SolMax717/1489.8%90.0%56.0%48.9%
8Claude Opus 4.8Max638/1480.0%56.1%52.0%65.6%
9Claude Haiku 5.5Max616/1475.1%46.3%55.0%78.9%
10GPT-5.5Xhigh566/1485.3%72.5%63.0%23.3%
11GPT-5.6 TerraMax557/1486.0%70.7%47.0%41.1%
12Gemini 3.7 FlashHigh497/1471.6%36.6%66.0%41.1%
13Qwen3.8-MaxXhigh458/1474.7%46.3%46.0%52.2%
14DeepSeek-V4.1-Flash記載なし448/1467.4%26.8%48.0%64.4%
15GPT-5.6 LunaMax387/1482.1%61.0%43.0%21.1%
15Kimi K3Max388/1472.2%39.0%36.0%54.4%
17Gemini 3.8 FlashHigh347/1468.4%22.0%58.0%38.9%
18GPT-6 LunaMax349/1479.0%56.1%36.0%26.7%
19Grok 4.6Xhigh247/1466.0%31.7%33.0%41.1%
20DeepSeek-V4-Pro-0813Max237/1464.6%26.8%46.0%33.3%
21Grok 4.7Xhigh217/1453.0%17.1%41.0%54.4%
22Claude Opus 4.6Max168/1466.0%26.8%21.0%33.3%
23Mistral Large 4High127/1458.6%26.8%33.0%27.8%
24Gemini 3.6 FlashHigh117/1459.0%22.0%40.0%23.3%
第三者・24 モデル・25 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)GPQA Diamond、OTIS Mock AIME 2024-2025、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(25)
  2. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(25)
  3. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(25)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. OTIS Mock AIME 2024-20253. MirrorCode
1Claude Sonnet 5.5Max927/1495.6%100.0%97.8%
2Gemini 4 ArgonHigh839/1496.4%100.0%58.9%
2GPT-6 AstraMax839/1495.8%100.0%60.0%
4GPT-6.1 TerraMax826/1494.3%100.0%76.7%
5GPT-6.1 SolMax819/1495.4%100.0%61.1%
6Claude Opus 5.5Max697/1490.6%100.0%95.6%
7GPT-6 SolMax677/1494.3%100.0%48.9%
8Gemini 3.8 FlashHigh547/1495.4%98.9%38.9%
9Claude Fable 5High534/1483.3%100.0%63.9%
9GPT-5.6 TerraMax537/1493.3%99.7%41.1%
11Claude Fable 5Max527/1485.9%99.7%70.0%
12Qwen3.8-MaxXhigh508/1492.7%99.4%52.2%
13Grok 4.6Xhigh497/1493.2%99.2%41.1%
14Claude Opus 4.8Max488/1491.0%98.3%65.6%
15Gemini 3.7 FlashHigh447/1494.8%97.2%41.1%
16DeepSeek-V4.1-Flash記載なし428/1489.8%98.3%64.4%
17Grok 4.7Xhigh417/1492.7%98.1%54.4%
18Kimi K3Max408/1493.1%97.2%54.4%
19DeepSeek-V4-Pro-0813Max347/1491.7%98.6%33.3%
20Gemini 3.1 Pro PreviewHigh286/1494.4%95.6%8.9%
21GPT-6 LunaMax279/1490.5%98.9%26.7%
22Gemini 3.6 FlashHigh257/1494.1%94.2%23.3%
23GPT-5.6 LunaMax237/1491.6%98.3%21.1%
24Mistral Large 4High197/1488.1%98.3%27.8%
25Claude Opus 4.6Max128/1488.4%91.1%33.3%
第三者・24 モデル・25 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)GPQA Diamond、Chess Puzzles(Epoch AI)、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(25)
  2. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(25)
  3. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(25)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. Chess Puzzles(Epoch AI)3. MirrorCode
1Claude Sonnet 5.5Max917/1495.6%59.0%97.8%
2GPT-6 AstraMax889/1495.8%74.0%60.0%
3GPT-6.1 SolMax849/1495.4%68.0%61.1%
4Gemini 4 ArgonHigh819/1496.4%59.0%58.9%
5Claude Opus 5.5Max727/1490.6%67.0%95.6%
6Gemini 3.7 FlashHigh687/1494.8%66.0%41.1%
7Gemini 3.8 FlashHigh637/1495.4%58.0%38.9%
8GPT-6 SolMax637/1494.3%56.0%48.9%
9Claude Opus 4.8Max608/1491.0%52.0%65.6%
10Claude Haiku 5.5Max586/1489.7%55.0%78.9%
11DeepSeek-V4.1-Flash記載なし518/1489.8%48.0%64.4%
12GPT-5.6 TerraMax497/1493.3%47.0%41.1%
13Claude Fable 5Max477/1485.9%47.0%70.0%
14Qwen3.8-MaxXhigh458/1492.7%46.0%52.2%
15Gemini 3.1 Pro PreviewHigh446/1494.4%51.0%8.9%
16Grok 4.7Xhigh447/1492.7%41.0%54.4%
17Kimi K3Max428/1493.1%36.0%54.4%
18Claude Fable 5High364/1483.3%45.0%63.9%
19DeepSeek-V4-Pro-0813Max357/1491.7%46.0%33.3%
20Grok 4.6Xhigh347/1493.2%33.0%41.1%
21Gemini 3.6 FlashHigh327/1494.1%40.0%23.3%
22GPT-5.6 LunaMax247/1491.6%43.0%21.1%
23GPT-6 LunaMax179/1490.5%36.0%26.7%
24Claude Opus 4.6Max128/1488.4%21.0%33.3%
25Mistral Large 4High107/1488.1%33.0%27.8%
第三者・24 モデル・24 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)GPQA Diamond、SimpleQA Verified、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(24)
  2. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(24)
  3. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(24)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. SimpleQA Verified3. MirrorCode
1Gemini 4 ArgonHigh889/1496.4%76.0%58.9%
2GPT-6 AstraMax879/1495.8%75.6%60.0%
3GPT-6.1 SolMax839/1495.4%73.9%61.1%
4Claude Sonnet 5.5Max747/1495.6%46.5%97.8%
5GPT-6.1 TerraMax726/1494.3%57.3%76.7%
6Claude Opus 5.5Max677/1490.6%72.2%95.6%
7Gemini 3.8 FlashHigh647/1495.4%69.7%38.9%
8Gemini 3.7 FlashHigh647/1494.8%69.2%41.1%
9GPT-6 SolMax607/1494.3%60.7%48.9%
10Gemini 3.1 Pro PreviewHigh546/1494.4%73.5%8.9%
10Claude Opus 4.8Max548/1491.0%53.0%65.6%
12Grok 4.7Xhigh537/1492.7%56.0%54.4%
13Kimi K3Max508/1493.1%50.6%54.4%
14Gemini 3.6 FlashHigh467/1494.1%66.2%23.3%
15Grok 4.6Xhigh437/1493.2%48.9%41.1%
16GPT-5.6 TerraMax397/1493.3%43.2%41.1%
16Qwen3.8-MaxXhigh398/1492.7%45.8%52.2%
18DeepSeek-V4.1-Flash記載なし368/1489.8%42.9%64.4%
19DeepSeek-V4-Pro-0813Max367/1491.7%52.9%33.3%
20Claude Haiku 5.5Max356/1489.7%23.8%78.9%
21Claude Opus 4.6Max218/1488.4%47.0%33.3%
22GPT-6 LunaMax149/1490.5%41.4%26.7%
22GPT-5.6 LunaMax147/1491.6%41.0%21.1%
24Mistral Large 4High67/1488.1%20.0%27.8%
第三者・24 モデル・24 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)SimpleQA Verified、Chess Puzzles(Epoch AI)、MirrorCode
  1. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(24)
  2. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(24)
  3. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(24)
順位モデルeffort平均百分位値のあるベンチマーク1. SimpleQA Verified2. Chess Puzzles(Epoch AI)3. MirrorCode
1Claude Opus 5.5Max907/1472.2%67.0%95.6%
1GPT-6 AstraMax909/1475.6%74.0%60.0%
3GPT-6.1 SolMax889/1473.9%68.0%61.1%
4Gemini 4 ArgonHigh829/1476.0%59.0%58.9%
5Claude Sonnet 5.5Max697/1446.5%59.0%97.8%
6Gemini 3.7 FlashHigh687/1469.2%66.0%41.1%
7Claude Opus 4.8Max658/1453.0%52.0%65.6%
8Gemini 3.8 FlashHigh617/1469.7%58.0%38.9%
9GPT-6 SolMax597/1460.7%56.0%48.9%
10GPT-5.5Xhigh536/1463.0%63.0%23.3%
11Claude Haiku 5.5Max526/1423.8%55.0%78.9%
12DeepSeek-V4.1-Flash記載なし498/1442.9%48.0%64.4%
13Grok 4.7Xhigh497/1456.0%41.0%54.4%
14Gemini 3.1 Pro PreviewHigh466/1473.5%51.0%8.9%
15Kimi K3Max418/1450.6%36.0%54.4%
16Qwen3.8-MaxXhigh408/1445.8%46.0%52.2%
17DeepSeek-V4-Pro-0813Max387/1452.9%46.0%33.3%
18GPT-5.6 TerraMax367/1443.2%47.0%41.1%
19Gemini 3.6 FlashHigh347/1466.2%40.0%23.3%
20Grok 4.6Xhigh307/1448.9%33.0%41.1%
21Claude Opus 4.6Max218/1447.0%21.0%33.3%
22GPT-6 LunaMax159/1441.4%36.0%26.7%
23GPT-5.6 LunaMax147/1441.0%43.0%21.1%
24Mistral Large 4High97/1420.0%33.0%27.8%
第三者・23 モデル・23 件 × 5 グループ(カバレッジ 5/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、OTIS Mock AIME 2024-2025、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(23)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(23)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(23)
  4. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(23)
  5. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(23)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. OTIS Mock AIME 2024-20255. MirrorCode
1GPT-6 AstraMax899/1495.8%93.7%97.6%100.0%60.0%
2GPT-6.1 SolMax889/1495.4%93.7%100.0%100.0%61.1%
3Claude Sonnet 5.5Max857/1495.6%88.8%80.5%100.0%97.8%
4GPT-6.1 TerraMax846/1494.3%90.9%92.7%100.0%76.7%
5Claude Opus 5.5Max777/1490.6%91.2%95.0%100.0%95.6%
6Gemini 4 ArgonHigh779/1496.4%86.0%78.1%100.0%58.9%
7GPT-6 SolMax727/1494.3%89.8%90.0%100.0%48.9%
8Claude Fable 5Max617/1485.9%87.0%90.2%99.7%70.0%
9GPT-5.6 TerraMax587/1493.3%86.0%70.7%99.7%41.1%
10Claude Opus 4.8Max498/1491.0%80.0%56.1%98.3%65.6%
11Qwen3.8-MaxXhigh488/1492.7%74.7%46.3%99.4%52.2%
12Kimi K3Max408/1493.1%72.2%39.0%97.2%54.4%
13Gemini 3.8 FlashHigh407/1495.4%68.4%22.0%98.9%38.9%
13Grok 4.6Xhigh407/1493.2%66.0%31.7%99.2%41.1%
13Gemini 3.7 FlashHigh407/1494.8%71.6%36.6%97.2%41.1%
16GPT-5.6 LunaMax367/1491.6%82.1%61.0%98.3%21.1%
17GPT-6 LunaMax359/1490.5%79.0%56.1%98.9%26.7%
18DeepSeek-V4.1-Flash記載なし348/1489.8%67.4%26.8%98.3%64.4%
19DeepSeek-V4-Pro-0813Max267/1491.7%64.6%26.8%98.6%33.3%
20Grok 4.7Xhigh247/1492.7%53.0%17.1%98.1%54.4%
21Gemini 3.6 FlashHigh187/1494.1%59.0%22.0%94.2%23.3%
22Mistral Large 4High157/1488.1%58.6%26.8%98.3%27.8%
23Claude Opus 4.6Max148/1488.4%66.0%26.8%91.1%33.3%
第三者・23 モデル・23 件 × 5 グループ(カバレッジ 5/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、SimpleQA Verified、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(23)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(23)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(23)
  4. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(23)
  5. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(23)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. SimpleQA Verified5. MirrorCode
1GPT-6 AstraMax909/1495.8%93.7%97.6%75.6%60.0%
2GPT-6.1 SolMax899/1495.4%93.7%100.0%73.9%61.1%
3Gemini 4 ArgonHigh819/1496.4%86.0%78.1%76.0%58.9%
4GPT-6.1 TerraMax796/1494.3%90.9%92.7%57.3%76.7%
5Claude Opus 5.5Max777/1490.6%91.2%95.0%72.2%95.6%
6Claude Sonnet 5.5Max757/1495.6%88.8%80.5%46.5%97.8%
7GPT-6 SolMax707/1494.3%89.8%90.0%60.7%48.9%
8Claude Opus 4.8Max568/1491.0%80.0%56.1%53.0%65.6%
9Gemini 3.7 FlashHigh537/1494.8%71.6%36.6%69.2%41.1%
10GPT-5.6 TerraMax517/1493.3%86.0%70.7%43.2%41.1%
11Kimi K3Max478/1493.1%72.2%39.0%50.6%54.4%
12Gemini 3.8 FlashHigh467/1495.4%68.4%22.0%69.7%38.9%
13Qwen3.8-MaxXhigh428/1492.7%74.7%46.3%45.8%52.2%
14Claude Haiku 5.5Max406/1489.7%75.1%46.3%23.8%78.9%
15Grok 4.6Xhigh377/1493.2%66.0%31.7%48.9%41.1%
16GPT-5.6 LunaMax347/1491.6%82.1%61.0%41.0%21.1%
17Grok 4.7Xhigh327/1492.7%53.0%17.1%56.0%54.4%
18Gemini 3.6 FlashHigh317/1494.1%59.0%22.0%66.2%23.3%
18DeepSeek-V4.1-Flash記載なし318/1489.8%67.4%26.8%42.9%64.4%
20GPT-6 LunaMax309/1490.5%79.0%56.1%41.4%26.7%
21DeepSeek-V4-Pro-0813Max287/1491.7%64.6%26.8%52.9%33.3%
22Claude Opus 4.6Max208/1488.4%66.0%26.8%47.0%33.3%
23Mistral Large 4High87/1488.1%58.6%26.8%20.0%27.8%
第三者・23 モデル・23 件 × 5 グループ(カバレッジ 5/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、Chess Puzzles(Epoch AI)、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(23)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(23)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(23)
  4. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(23)
  5. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(23)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. Chess Puzzles(Epoch AI)5. MirrorCode
1GPT-6 AstraMax919/1495.8%93.7%97.6%74.0%60.0%
2GPT-6.1 SolMax909/1495.4%93.7%100.0%68.0%61.1%
3Claude Sonnet 5.5Max867/1495.6%88.8%80.5%59.0%97.8%
4Claude Opus 5.5Max797/1490.6%91.2%95.0%67.0%95.6%
5Gemini 4 ArgonHigh779/1496.4%86.0%78.1%59.0%58.9%
6GPT-6 SolMax717/1494.3%89.8%90.0%56.0%48.9%
7Claude Fable 5Max607/1485.9%87.0%90.2%47.0%70.0%
8Claude Opus 4.8Max588/1491.0%80.0%56.1%52.0%65.6%
9GPT-5.6 TerraMax577/1493.3%86.0%70.7%47.0%41.1%
10Gemini 3.7 FlashHigh557/1494.8%71.6%36.6%66.0%41.1%
11Claude Haiku 5.5Max536/1489.7%75.1%46.3%55.0%78.9%
12Qwen3.8-MaxXhigh458/1492.7%74.7%46.3%46.0%52.2%
13Gemini 3.8 FlashHigh457/1495.4%68.4%22.0%58.0%38.9%
14Kimi K3Max428/1493.1%72.2%39.0%36.0%54.4%
15DeepSeek-V4.1-Flash記載なし408/1489.8%67.4%26.8%48.0%64.4%
16GPT-5.6 LunaMax397/1491.6%82.1%61.0%43.0%21.1%
17GPT-6 LunaMax329/1490.5%79.0%56.1%36.0%26.7%
18Grok 4.6Xhigh317/1493.2%66.0%31.7%33.0%41.1%
19Grok 4.7Xhigh277/1492.7%53.0%17.1%41.0%54.4%
19DeepSeek-V4-Pro-0813Max277/1491.7%64.6%26.8%46.0%33.3%
21Gemini 3.6 FlashHigh227/1494.1%59.0%22.0%40.0%23.3%
22Claude Opus 4.6Max148/1488.4%66.0%26.8%21.0%33.3%
23Mistral Large 4High107/1488.1%58.6%26.8%33.0%27.8%
第三者・23 モデル・23 件 × 5 グループ(カバレッジ 5/14 ベンチマーク)FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、SimpleQA Verified、Chess Puzzles(Epoch AI)、MirrorCode
  1. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(23)
  2. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(23)
  3. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(23)
  4. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(23)
  5. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(23)
順位モデルeffort平均百分位値のあるベンチマーク1. FrontierMath Tiers 1-3(v2)2. FrontierMath Tier 4(v2)3. SimpleQA Verified4. Chess Puzzles(Epoch AI)5. MirrorCode
1GPT-6.1 SolMax929/1493.7%100.0%73.9%68.0%61.1%
1GPT-6 AstraMax929/1493.7%97.6%75.6%74.0%60.0%
3Claude Opus 5.5Max917/1491.2%95.0%72.2%67.0%95.6%
4Gemini 4 ArgonHigh799/1486.0%78.1%76.0%59.0%58.9%
5Claude Sonnet 5.5Max747/1488.8%80.5%46.5%59.0%97.8%
6GPT-6 SolMax707/1489.8%90.0%60.7%56.0%48.9%
7Claude Opus 4.8Max628/1480.0%56.1%53.0%52.0%65.6%
8GPT-5.5Xhigh606/1485.3%72.5%63.0%63.0%23.3%
9Gemini 3.7 FlashHigh557/1471.6%36.6%69.2%66.0%41.1%
10Claude Haiku 5.5Max506/1475.1%46.3%23.8%55.0%78.9%
10GPT-5.6 TerraMax507/1486.0%70.7%43.2%47.0%41.1%
12Gemini 3.8 FlashHigh447/1468.4%22.0%69.7%58.0%38.9%
13Qwen3.8-MaxXhigh438/1474.7%46.3%45.8%46.0%52.2%
14Kimi K3Max418/1472.2%39.0%50.6%36.0%54.4%
15DeepSeek-V4.1-Flash記載なし408/1467.4%26.8%42.9%48.0%64.4%
16GPT-5.6 LunaMax347/1482.1%61.0%41.0%43.0%21.1%
17GPT-6 LunaMax319/1479.0%56.1%41.4%36.0%26.7%
18Grok 4.7Xhigh307/1453.0%17.1%56.0%41.0%54.4%
18DeepSeek-V4-Pro-0813Max307/1464.6%26.8%52.9%46.0%33.3%
20Grok 4.6Xhigh287/1466.0%31.7%48.9%33.0%41.1%
21Gemini 3.6 FlashHigh247/1459.0%22.0%66.2%40.0%23.3%
22Claude Opus 4.6Max208/1466.0%26.8%47.0%21.0%33.3%
23Mistral Large 4High107/1458.6%26.8%20.0%33.0%27.8%
第三者・23 モデル・24 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)GPQA Diamond、OTIS Mock AIME 2024-2025、Chess Puzzles(Epoch AI)、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(24)
  2. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(24)
  3. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(24)
  4. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(24)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. OTIS Mock AIME 2024-20253. Chess Puzzles(Epoch AI)4. MirrorCode
1Claude Sonnet 5.5Max907/1495.6%100.0%59.0%97.8%
2GPT-6 AstraMax889/1495.8%100.0%74.0%60.0%
3GPT-6.1 SolMax859/1495.4%100.0%68.0%61.1%
4Gemini 4 ArgonHigh839/1496.4%100.0%59.0%58.9%
5Claude Opus 5.5Max757/1490.6%100.0%67.0%95.6%
6GPT-6 SolMax687/1494.3%100.0%56.0%48.9%
7Gemini 3.8 FlashHigh607/1495.4%98.9%58.0%38.9%
8Gemini 3.7 FlashHigh557/1494.8%97.2%66.0%41.1%
9GPT-5.6 TerraMax547/1493.3%99.7%47.0%41.1%
10Claude Opus 4.8Max548/1491.0%98.3%52.0%65.6%
11Claude Fable 5Max537/1485.9%99.7%47.0%70.0%
12Claude Fable 5High504/1483.3%100.0%45.0%63.9%
13Qwen3.8-MaxXhigh498/1492.7%99.4%46.0%52.2%
14DeepSeek-V4.1-Flash記載なし478/1489.8%98.3%48.0%64.4%
15Grok 4.6Xhigh407/1493.2%99.2%33.0%41.1%
16Grok 4.7Xhigh397/1492.7%98.1%41.0%54.4%
17DeepSeek-V4-Pro-0813Max377/1491.7%98.6%46.0%33.3%
18Gemini 3.1 Pro PreviewHigh366/1494.4%95.6%51.0%8.9%
19Kimi K3Max358/1493.1%97.2%36.0%54.4%
20GPT-5.6 LunaMax267/1491.6%98.3%43.0%21.1%
21GPT-6 LunaMax259/1490.5%98.9%36.0%26.7%
21Gemini 3.6 FlashHigh257/1494.1%94.2%40.0%23.3%
23Mistral Large 4High167/1488.1%98.3%33.0%27.8%
24Claude Opus 4.6Max98/1488.4%91.1%21.0%33.3%
第三者・23 モデル・23 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)GPQA Diamond、OTIS Mock AIME 2024-2025、SimpleQA Verified、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(23)
  2. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(23)
  3. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(23)
  4. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(23)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. OTIS Mock AIME 2024-20253. SimpleQA Verified4. MirrorCode
1Gemini 4 ArgonHigh899/1496.4%100.0%76.0%58.9%
2GPT-6 AstraMax889/1495.8%100.0%75.6%60.0%
3GPT-6.1 SolMax859/1495.4%100.0%73.9%61.1%
4Claude Sonnet 5.5Max767/1495.6%100.0%46.5%97.8%
5GPT-6.1 TerraMax766/1494.3%100.0%57.3%76.7%
6Claude Opus 5.5Max707/1490.6%100.0%72.2%95.6%
7GPT-6 SolMax667/1494.3%100.0%60.7%48.9%
8Gemini 3.8 FlashHigh617/1495.4%98.9%69.7%38.9%
9Gemini 3.7 FlashHigh527/1494.8%97.2%69.2%41.1%
10Claude Opus 4.8Max488/1491.0%98.3%53.0%65.6%
11Grok 4.6Xhigh477/1493.2%99.2%48.9%41.1%
12GPT-5.6 TerraMax457/1493.3%99.7%43.2%41.1%
13Grok 4.7Xhigh457/1492.7%98.1%56.0%54.4%
14Qwen3.8-MaxXhigh448/1492.7%99.4%45.8%52.2%
15Gemini 3.1 Pro PreviewHigh426/1494.4%95.6%73.5%8.9%
16Kimi K3Max418/1493.1%97.2%50.6%54.4%
17DeepSeek-V4-Pro-0813Max377/1491.7%98.6%52.9%33.3%
18Gemini 3.6 FlashHigh357/1494.1%94.2%66.2%23.3%
19DeepSeek-V4.1-Flash記載なし358/1489.8%98.3%42.9%64.4%
20GPT-6 LunaMax229/1490.5%98.9%41.4%26.7%
21GPT-5.6 LunaMax187/1491.6%98.3%41.0%21.1%
22Claude Opus 4.6Max158/1488.4%91.1%47.0%33.3%
23Mistral Large 4High137/1488.1%98.3%20.0%27.8%
第三者・23 モデル・23 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)GPQA Diamond、SimpleQA Verified、Chess Puzzles(Epoch AI)、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(23)
  2. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(23)
  3. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(23)
  4. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(23)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. SimpleQA Verified3. Chess Puzzles(Epoch AI)4. MirrorCode
1GPT-6 AstraMax919/1495.8%75.6%74.0%60.0%
2Gemini 4 ArgonHigh879/1496.4%76.0%59.0%58.9%
2GPT-6.1 SolMax879/1495.4%73.9%68.0%61.1%
4Claude Sonnet 5.5Max767/1495.6%46.5%59.0%97.8%
5Claude Opus 5.5Max737/1490.6%72.2%67.0%95.6%
6Gemini 3.7 FlashHigh697/1494.8%69.2%66.0%41.1%
7Gemini 3.8 FlashHigh667/1495.4%69.7%58.0%38.9%
8GPT-6 SolMax637/1494.3%60.7%56.0%48.9%
9Claude Opus 4.8Max578/1491.0%53.0%52.0%65.6%
10Gemini 3.1 Pro PreviewHigh536/1494.4%73.5%51.0%8.9%
11Grok 4.7Xhigh487/1492.7%56.0%41.0%54.4%
12Kimi K3Max438/1493.1%50.6%36.0%54.4%
13Claude Haiku 5.5Max426/1489.7%23.8%55.0%78.9%
13GPT-5.6 TerraMax427/1493.3%43.2%47.0%41.1%
15Gemini 3.6 FlashHigh417/1494.1%66.2%40.0%23.3%
15DeepSeek-V4.1-Flash記載なし418/1489.8%42.9%48.0%64.4%
17Qwen3.8-MaxXhigh408/1492.7%45.8%46.0%52.2%
18DeepSeek-V4-Pro-0813Max387/1491.7%52.9%46.0%33.3%
19Grok 4.6Xhigh367/1493.2%48.9%33.0%41.1%
20GPT-5.6 LunaMax197/1491.6%41.0%43.0%21.1%
21Claude Opus 4.6Max168/1488.4%47.0%21.0%33.3%
22GPT-6 LunaMax159/1490.5%41.4%36.0%26.7%
23Mistral Large 4High67/1488.1%20.0%33.0%27.8%
第三者・22 モデル・22 件 × 6 グループ(カバレッジ 6/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、OTIS Mock AIME 2024-2025、SimpleQA Verified、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(22)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(22)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(22)
  4. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(22)
  5. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(22)
  6. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(22)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. OTIS Mock AIME 2024-20255. SimpleQA Verified6. MirrorCode
1GPT-6 AstraMax909/1495.8%93.7%97.6%100.0%75.6%60.0%
2GPT-6.1 SolMax899/1495.4%93.7%100.0%100.0%73.9%61.1%
3Gemini 4 ArgonHigh829/1496.4%86.0%78.1%100.0%76.0%58.9%
4GPT-6.1 TerraMax806/1494.3%90.9%92.7%100.0%57.3%76.7%
5Claude Opus 5.5Max787/1490.6%91.2%95.0%100.0%72.2%95.6%
6Claude Sonnet 5.5Max767/1495.6%88.8%80.5%100.0%46.5%97.8%
7GPT-6 SolMax727/1494.3%89.8%90.0%100.0%60.7%48.9%
8GPT-5.6 TerraMax537/1493.3%86.0%70.7%99.7%43.2%41.1%
9Claude Opus 4.8Max518/1491.0%80.0%56.1%98.3%53.0%65.6%
10Gemini 3.8 FlashHigh477/1495.4%68.4%22.0%98.9%69.7%38.9%
11Gemini 3.7 FlashHigh467/1494.8%71.6%36.6%97.2%69.2%41.1%
12Qwen3.8-MaxXhigh458/1492.7%74.7%46.3%99.4%45.8%52.2%
13Kimi K3Max418/1493.1%72.2%39.0%97.2%50.6%54.4%
14Grok 4.6Xhigh407/1493.2%66.0%31.7%99.2%48.9%41.1%
15GPT-6 LunaMax329/1490.5%79.0%56.1%98.9%41.4%26.7%
16GPT-5.6 LunaMax317/1491.6%82.1%61.0%98.3%41.0%21.1%
17DeepSeek-V4.1-Flash記載なし318/1489.8%67.4%26.8%98.3%42.9%64.4%
18DeepSeek-V4-Pro-0813Max307/1491.7%64.6%26.8%98.6%52.9%33.3%
19Grok 4.7Xhigh307/1492.7%53.0%17.1%98.1%56.0%54.4%
20Gemini 3.6 FlashHigh277/1494.1%59.0%22.0%94.2%66.2%23.3%
21Claude Opus 4.6Max178/1488.4%66.0%26.8%91.1%47.0%33.3%
22Mistral Large 4High127/1488.1%58.6%26.8%98.3%20.0%27.8%
第三者・22 モデル・22 件 × 6 グループ(カバレッジ 6/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、OTIS Mock AIME 2024-2025、Chess Puzzles(Epoch AI)、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(22)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(22)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(22)
  4. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(22)
  5. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(22)
  6. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(22)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. OTIS Mock AIME 2024-20255. Chess Puzzles(Epoch AI)6. MirrorCode
1GPT-6 AstraMax919/1495.8%93.7%97.6%100.0%74.0%60.0%
2GPT-6.1 SolMax909/1495.4%93.7%100.0%100.0%68.0%61.1%
3Claude Sonnet 5.5Max867/1495.6%88.8%80.5%100.0%59.0%97.8%
4Claude Opus 5.5Max807/1490.6%91.2%95.0%100.0%67.0%95.6%
5Gemini 4 ArgonHigh799/1496.4%86.0%78.1%100.0%59.0%58.9%
6GPT-6 SolMax737/1494.3%89.8%90.0%100.0%56.0%48.9%
7Claude Fable 5Max627/1485.9%87.0%90.2%99.7%47.0%70.0%
8GPT-5.6 TerraMax597/1493.3%86.0%70.7%99.7%47.0%41.1%
9Claude Opus 4.8Max538/1491.0%80.0%56.1%98.3%52.0%65.6%
10Qwen3.8-MaxXhigh498/1492.7%74.7%46.3%99.4%46.0%52.2%
11Gemini 3.7 FlashHigh487/1494.8%71.6%36.6%97.2%66.0%41.1%
12Gemini 3.8 FlashHigh467/1495.4%68.4%22.0%98.9%58.0%38.9%
13DeepSeek-V4.1-Flash記載なし398/1489.8%67.4%26.8%98.3%48.0%64.4%
14Kimi K3Max388/1493.1%72.2%39.0%97.2%36.0%54.4%
15GPT-5.6 LunaMax377/1491.6%82.1%61.0%98.3%43.0%21.1%
16Grok 4.6Xhigh367/1493.2%66.0%31.7%99.2%33.0%41.1%
17GPT-6 LunaMax349/1490.5%79.0%56.1%98.9%36.0%26.7%
18DeepSeek-V4-Pro-0813Max307/1491.7%64.6%26.8%98.6%46.0%33.3%
19Grok 4.7Xhigh267/1492.7%53.0%17.1%98.1%41.0%54.4%
20Gemini 3.6 FlashHigh197/1494.1%59.0%22.0%94.2%40.0%23.3%
21Mistral Large 4High147/1488.1%58.6%26.8%98.3%33.0%27.8%
22Claude Opus 4.6Max128/1488.4%66.0%26.8%91.1%21.0%33.3%
第三者・22 モデル・22 件 × 6 グループ(カバレッジ 6/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、SimpleQA Verified、Chess Puzzles(Epoch AI)、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(22)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(22)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(22)
  4. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(22)
  5. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(22)
  6. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(22)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. SimpleQA Verified5. Chess Puzzles(Epoch AI)6. MirrorCode
1GPT-6 AstraMax929/1495.8%93.7%97.6%75.6%74.0%60.0%
2GPT-6.1 SolMax909/1495.4%93.7%100.0%73.9%68.0%61.1%
3Gemini 4 ArgonHigh839/1496.4%86.0%78.1%76.0%59.0%58.9%
4Claude Opus 5.5Max797/1490.6%91.2%95.0%72.2%67.0%95.6%
5Claude Sonnet 5.5Max777/1495.6%88.8%80.5%46.5%59.0%97.8%
6GPT-6 SolMax717/1494.3%89.8%90.0%60.7%56.0%48.9%
7Gemini 3.7 FlashHigh597/1494.8%71.6%36.6%69.2%66.0%41.1%
8Claude Opus 4.8Max588/1491.0%80.0%56.1%53.0%52.0%65.6%
9GPT-5.6 TerraMax537/1493.3%86.0%70.7%43.2%47.0%41.1%
10Gemini 3.8 FlashHigh517/1495.4%68.4%22.0%69.7%58.0%38.9%
11Claude Haiku 5.5Max456/1489.7%75.1%46.3%23.8%55.0%78.9%
12Kimi K3Max448/1493.1%72.2%39.0%50.6%36.0%54.4%
12Qwen3.8-MaxXhigh448/1492.7%74.7%46.3%45.8%46.0%52.2%
14DeepSeek-V4.1-Flash記載なし368/1489.8%67.4%26.8%42.9%48.0%64.4%
15GPT-5.6 LunaMax357/1491.6%82.1%61.0%41.0%43.0%21.1%
16Grok 4.6Xhigh337/1493.2%66.0%31.7%48.9%33.0%41.1%
17Grok 4.7Xhigh337/1492.7%53.0%17.1%56.0%41.0%54.4%
18DeepSeek-V4-Pro-0813Max317/1491.7%64.6%26.8%52.9%46.0%33.3%
19Gemini 3.6 FlashHigh317/1494.1%59.0%22.0%66.2%40.0%23.3%
20GPT-6 LunaMax299/1490.5%79.0%56.1%41.4%36.0%26.7%
21Claude Opus 4.6Max188/1488.4%66.0%26.8%47.0%21.0%33.3%
22Mistral Large 4High87/1488.1%58.6%26.8%20.0%33.0%27.8%
第三者・22 モデル・22 件 × 5 グループ(カバレッジ 5/14 ベンチマーク)GPQA Diamond、OTIS Mock AIME 2024-2025、SimpleQA Verified、Chess Puzzles(Epoch AI)、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(22)
  2. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(22)
  3. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(22)
  4. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(22)
  5. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(22)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. OTIS Mock AIME 2024-20253. SimpleQA Verified4. Chess Puzzles(Epoch AI)5. MirrorCode
1GPT-6 AstraMax919/1495.8%100.0%75.6%74.0%60.0%
2Gemini 4 ArgonHigh889/1496.4%100.0%76.0%59.0%58.9%
2GPT-6.1 SolMax889/1495.4%100.0%73.9%68.0%61.1%
4Claude Sonnet 5.5Max777/1495.6%100.0%46.5%59.0%97.8%
5Claude Opus 5.5Max757/1490.6%100.0%72.2%67.0%95.6%
6GPT-6 SolMax677/1494.3%100.0%60.7%56.0%48.9%
7Gemini 3.8 FlashHigh647/1495.4%98.9%69.7%58.0%38.9%
8Gemini 3.7 FlashHigh597/1494.8%97.2%69.2%66.0%41.1%
9Claude Opus 4.8Max538/1491.0%98.3%53.0%52.0%65.6%
10GPT-5.6 TerraMax487/1493.3%99.7%43.2%47.0%41.1%
11Qwen3.8-MaxXhigh458/1492.7%99.4%45.8%46.0%52.2%
12Gemini 3.1 Pro PreviewHigh456/1494.4%95.6%73.5%51.0%8.9%
13Grok 4.7Xhigh437/1492.7%98.1%56.0%41.0%54.4%
14Grok 4.6Xhigh407/1493.2%99.2%48.9%33.0%41.1%
15DeepSeek-V4.1-Flash記載なし408/1489.8%98.3%42.9%48.0%64.4%
16DeepSeek-V4-Pro-0813Max397/1491.7%98.6%52.9%46.0%33.3%
17Kimi K3Max388/1493.1%97.2%50.6%36.0%54.4%
18Gemini 3.6 FlashHigh337/1494.1%94.2%66.2%40.0%23.3%
19GPT-6 LunaMax229/1490.5%98.9%41.4%36.0%26.7%
20GPT-5.6 LunaMax217/1491.6%98.3%41.0%43.0%21.1%
21Claude Opus 4.6Max138/1488.4%91.1%47.0%21.0%33.3%
22Mistral Large 4High127/1488.1%98.3%20.0%33.0%27.8%
第三者・21 モデル・21 件 × 7 グループ(カバレッジ 7/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、OTIS Mock AIME 2024-2025、SimpleQA Verified、Chess Puzzles(Epoch AI)、MirrorCode
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(21)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(21)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(21)
  4. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(21)
  5. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(21)
  6. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(21)
  7. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(21)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. OTIS Mock AIME 2024-20255. SimpleQA Verified6. Chess Puzzles(Epoch AI)7. MirrorCode
1GPT-6 AstraMax929/1495.8%93.7%97.6%100.0%75.6%74.0%60.0%
2GPT-6.1 SolMax909/1495.4%93.7%100.0%100.0%73.9%68.0%61.1%
3Gemini 4 ArgonHigh839/1496.4%86.0%78.1%100.0%76.0%59.0%58.9%
4Claude Opus 5.5Max807/1490.6%91.2%95.0%100.0%72.2%67.0%95.6%
5Claude Sonnet 5.5Max787/1495.6%88.8%80.5%100.0%46.5%59.0%97.8%
6GPT-6 SolMax737/1494.3%89.8%90.0%100.0%60.7%56.0%48.9%
7GPT-5.6 TerraMax557/1493.3%86.0%70.7%99.7%43.2%47.0%41.1%
8Claude Opus 4.8Max548/1491.0%80.0%56.1%98.3%53.0%52.0%65.6%
9Gemini 3.7 FlashHigh537/1494.8%71.6%36.6%97.2%69.2%66.0%41.1%
10Gemini 3.8 FlashHigh517/1495.4%68.4%22.0%98.9%69.7%58.0%38.9%
11Qwen3.8-MaxXhigh478/1492.7%74.7%46.3%99.4%45.8%46.0%52.2%
12Kimi K3Max408/1493.1%72.2%39.0%97.2%50.6%36.0%54.4%
13Grok 4.6Xhigh377/1493.2%66.0%31.7%99.2%48.9%33.0%41.1%
14DeepSeek-V4.1-Flash記載なし368/1489.8%67.4%26.8%98.3%42.9%48.0%64.4%
15GPT-5.6 LunaMax337/1491.6%82.1%61.0%98.3%41.0%43.0%21.1%
15DeepSeek-V4-Pro-0813Max337/1491.7%64.6%26.8%98.6%52.9%46.0%33.3%
17GPT-6 LunaMax319/1490.5%79.0%56.1%98.9%41.4%36.0%26.7%
17Grok 4.7Xhigh317/1492.7%53.0%17.1%98.1%56.0%41.0%54.4%
19Gemini 3.6 FlashHigh277/1494.1%59.0%22.0%94.2%66.2%40.0%23.3%
20Claude Opus 4.6Max158/1488.4%66.0%26.8%91.1%47.0%21.0%33.3%
21Mistral Large 4High127/1488.1%58.6%26.8%98.3%20.0%33.0%27.8%
第三者・7 モデル・7 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)GPQA Diamond、OTIS Mock AIME 2024-2025、Chess Puzzles(Epoch AI)、τ²-bench banking_knowledge (v1.0.1)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(7)
  2. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(7)
  3. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(7)
  4. τ²-bench banking_knowledge (v1.0.1)第三者・τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 alltools)計測・報告: Sierra(7)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. OTIS Mock AIME 2024-20253. Chess Puzzles(Epoch AI)4. τ²-bench banking_knowledge (v1.0.1)
1Qwen3.8-MaxXhigh758/1492.7%99.4%46.0%55.2%
2Claude Opus 4.8Max698/1491.0%98.3%52.0%39.7%
3Claude Fable 5Max567/1485.9%99.7%47.0%39.7%
4Gemini 3.1 Pro PreviewHigh546/1494.4%95.6%51.0%26.0%
5Kimi K3Max468/1493.1%97.2%36.0%37.1%
6Claude Opus 4.7Max336/1486.4%86.7%41.0%40.2%
7Claude Opus 4.6Max178/1488.4%91.1%21.0%27.3%
第三者・7 モデル・7 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、Chess Puzzles(Epoch AI)、τ²-bench banking_knowledge (v1.0.1)
  1. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(7)
  2. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(7)
  3. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(7)
  4. τ²-bench banking_knowledge (v1.0.1)第三者・τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 alltools)計測・報告: Sierra(7)
順位モデルeffort平均百分位値のあるベンチマーク1. FrontierMath Tiers 1-3(v2)2. FrontierMath Tier 4(v2)3. Chess Puzzles(Epoch AI)4. τ²-bench banking_knowledge (v1.0.1)
1GPT-5.5Xhigh886/1485.3%72.5%63.0%44.6%
2Claude Fable 5Max777/1487.0%90.2%47.0%39.7%
3Claude Opus 4.8Max658/1480.0%56.1%52.0%39.7%
4Qwen3.8-MaxXhigh638/1474.7%46.3%46.0%55.2%
5Claude Opus 4.7Max336/1470.2%31.7%41.0%40.2%
6Kimi K3Max258/1472.2%39.0%36.0%37.1%
7Claude Opus 4.6Max08/1466.0%26.8%21.0%27.3%
第三者・7 モデル・8 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)τ²-bench airline (v1.0.1)、τ²-bench retail (v1.0.1)、τ²-bench telecom (v1.0.1)
  1. τ²-bench airline (v1.0.1)第三者・τ²-bench v1.0.1(Sierra の実行)計測・報告: Sierra(8)
  2. τ²-bench retail (v1.0.1)第三者・τ²-bench v1.0.1(Sierra の実行)計測・報告: Sierra(8)
  3. τ²-bench telecom (v1.0.1)第三者・τ²-bench v1.0.1(Sierra の実行)計測・報告: Sierra(8)
順位モデルeffort平均百分位値のあるベンチマーク1. τ²-bench airline (v1.0.1)2. τ²-bench retail (v1.0.1)3. τ²-bench telecom (v1.0.1)
1Claude Opus 4.5High864/1484.0%79.6%92.3%
2Qwen3.5-397B-A17B記載なし814/1481.5%84.4%97.8%
3GPT-5.2High714/1483.0%81.6%89.7%
4Gemini 3 Flash PreviewHigh648/1482.5%76.8%91.2%
5Gemini 3 Pro PreviewHigh434/1480.5%75.9%91.0%
6GLM-5記載なし364/1482.5%73.7%86.8%
7Claude Sonnet 4.5記載なし109/1472.0%72.4%84.9%
7GPT-5.2None104/1452.5%75.0%57.2%
第三者・7 モデル・7 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)GPQA Diamond、OTIS Mock AIME 2024-2025、Epoch Capabilities Index(ECI)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(7)
  2. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(7)
  3. Epoch Capabilities Index(ECI)第三者・Epoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)計測・報告: Epoch AI(7)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. OTIS Mock AIME 2024-20253. Epoch Capabilities Index(ECI)
1DeepSeek-V4.1-Flash記載なし898/1489.8%98.3%156.7
2Gemini 3.1 Pro Preview記載なし835/1494.1%95.6%156.18
3Claude Opus 4.8記載なし674/1485.4%84.4%159.94
4Gemini 3 Flash Preview記載なし505/1483.2%92.8%152.32
5Gemini 2.5 Pro記載なし335/1485.3%84.7%144.87
6Claude Opus 4.5記載なし224/1480.7%48.1%151.92
7Claude Sonnet 4.5記載なし69/1473.7%35.6%147.63
第三者・7 モデル・7 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)Chess Puzzles(Epoch AI)、MirrorCode、τ²-bench banking_knowledge (v1.0.1)
  1. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(7)
  2. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(7)
  3. τ²-bench banking_knowledge (v1.0.1)第三者・τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 alltools)計測・報告: Sierra(7)
順位モデルeffort平均百分位値のあるベンチマーク1. Chess Puzzles(Epoch AI)2. MirrorCode3. τ²-bench banking_knowledge (v1.0.1)
1Claude Opus 4.8Max758/1452.0%65.6%39.7%
2Claude Fable 5Max697/1447.0%70.0%39.7%
3GPT-5.5Xhigh676/1463.0%23.3%44.6%
4Qwen3.8-MaxXhigh618/1446.0%52.2%55.2%
5Kimi K3Max398/1436.0%54.4%37.1%
6Gemini 3.1 Pro PreviewHigh226/1451.0%8.9%26.0%
7Claude Opus 4.6Max178/1421.0%33.3%27.3%
第三者・6 モデル・6 件 × 6 グループ(カバレッジ 6/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、OTIS Mock AIME 2024-2025、Chess Puzzles(Epoch AI)、τ²-bench banking_knowledge (v1.0.1)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(6)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(6)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(6)
  4. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(6)
  5. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(6)
  6. τ²-bench banking_knowledge (v1.0.1)第三者・τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 alltools)計測・報告: Sierra(6)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. OTIS Mock AIME 2024-20255. Chess Puzzles(Epoch AI)6. τ²-bench banking_knowledge (v1.0.1)
1Qwen3.8-MaxXhigh738/1492.7%74.7%46.3%99.4%46.0%55.2%
2Claude Fable 5Max727/1485.9%87.0%90.2%99.7%47.0%39.7%
2Claude Opus 4.8Max728/1491.0%80.0%56.1%98.3%52.0%39.7%
4Kimi K3Max438/1493.1%72.2%39.0%97.2%36.0%37.1%
5Claude Opus 4.7Max306/1486.4%70.2%31.7%86.7%41.0%40.2%
6Claude Opus 4.6Max108/1488.4%66.0%26.8%91.1%21.0%27.3%
第三者・6 モデル・6 件 × 5 グループ(カバレッジ 5/14 ベンチマーク)GPQA Diamond、OTIS Mock AIME 2024-2025、Chess Puzzles(Epoch AI)、MirrorCode、τ²-bench banking_knowledge (v1.0.1)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(6)
  2. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(6)
  3. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(6)
  4. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(6)
  5. τ²-bench banking_knowledge (v1.0.1)第三者・τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 alltools)計測・報告: Sierra(6)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. OTIS Mock AIME 2024-20253. Chess Puzzles(Epoch AI)4. MirrorCode5. τ²-bench banking_knowledge (v1.0.1)
1Claude Opus 4.8Max708/1491.0%98.3%52.0%65.6%39.7%
2Claude Fable 5Max667/1485.9%99.7%47.0%70.0%39.7%
3Qwen3.8-MaxXhigh648/1492.7%99.4%46.0%52.2%55.2%
4Kimi K3Max488/1493.1%97.2%36.0%54.4%37.1%
5Gemini 3.1 Pro PreviewHigh406/1494.4%95.6%51.0%8.9%26.0%
6Claude Opus 4.6Max128/1488.4%91.1%21.0%33.3%27.3%
第三者・6 モデル・6 件 × 5 グループ(カバレッジ 5/14 ベンチマーク)FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、Chess Puzzles(Epoch AI)、MirrorCode、τ²-bench banking_knowledge (v1.0.1)
  1. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(6)
  2. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(6)
  3. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(6)
  4. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(6)
  5. τ²-bench banking_knowledge (v1.0.1)第三者・τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 alltools)計測・報告: Sierra(6)
順位モデルeffort平均百分位値のあるベンチマーク1. FrontierMath Tiers 1-3(v2)2. FrontierMath Tier 4(v2)3. Chess Puzzles(Epoch AI)4. MirrorCode5. τ²-bench banking_knowledge (v1.0.1)
1Claude Fable 5Max827/1487.0%90.2%47.0%70.0%39.7%
2GPT-5.5Xhigh686/1485.3%72.5%63.0%23.3%44.6%
3Claude Opus 4.8Max668/1480.0%56.1%52.0%65.6%39.7%
4Qwen3.8-MaxXhigh528/1474.7%46.3%46.0%52.2%55.2%
5Kimi K3Max288/1472.2%39.0%36.0%54.4%37.1%
6Claude Opus 4.6Max48/1466.0%26.8%21.0%33.3%27.3%
第三者・6 モデル・6 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)SimpleQA Verified、Chess Puzzles(Epoch AI)、MirrorCode、τ²-bench banking_knowledge (v1.0.1)
  1. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(6)
  2. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(6)
  3. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(6)
  4. τ²-bench banking_knowledge (v1.0.1)第三者・τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 alltools)計測・報告: Sierra(6)
順位モデルeffort平均百分位値のあるベンチマーク1. SimpleQA Verified2. Chess Puzzles(Epoch AI)3. MirrorCode4. τ²-bench banking_knowledge (v1.0.1)
1Claude Opus 4.8Max758/1453.0%52.0%65.6%39.7%
2GPT-5.5Xhigh706/1463.0%63.0%23.3%44.6%
3Qwen3.8-MaxXhigh508/1445.8%46.0%52.2%55.2%
4Kimi K3Max458/1450.6%36.0%54.4%37.1%
5Gemini 3.1 Pro PreviewHigh406/1473.5%51.0%8.9%26.0%
6Claude Opus 4.6Max208/1447.0%21.0%33.3%27.3%
第三者・6 モデル・6 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)OTIS Mock AIME 2024-2025、LMArena Text(style control・overall)、LMArena Text Coding(style control なし)
  1. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(6)
  2. LMArena Text(style control・overall)第三者・lmarena-ai/leaderboard-dataset text_style_control/latest-00000-of-00001.parquet (CC BY 4.0)計測・報告: LMArena(6)
  3. LMArena Text Coding(style control なし)第三者・lmarena-ai/leaderboard-dataset text/latest-00000-of-00001.parquet (CC BY 4.0)計測・報告: LMArena(6)
順位モデルeffort平均百分位値のあるベンチマーク1. OTIS Mock AIME 2024-20252. LMArena Text(style control・overall)3. LMArena Text Coding(style control なし)
1Gemini 4 CarbonHigh908/14100.0%1,5261,563
2Gemini 4 ArgonHigh779/14100.0%1,5221,540
3GPT-6.1 SolMax579/14100.0%1,4821,486
3GPT-6 AstraMax579/14100.0%1,4741,494
5GPT-6 LunaMax139/1498.9%1,4411,432
6Claude Haiku 5.5High73/1497.2%1,4221,454
第三者・5 モデル・5 件 × 8 グループ(カバレッジ 8/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、OTIS Mock AIME 2024-2025、SimpleQA Verified、Chess Puzzles(Epoch AI)、LMArena Text(style control・overall)、LMArena Text Coding(style control なし)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(5)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(5)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(5)
  4. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(5)
  5. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(5)
  6. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(5)
  7. LMArena Text(style control・overall)第三者・lmarena-ai/leaderboard-dataset text_style_control/latest-00000-of-00001.parquet (CC BY 4.0)計測・報告: LMArena(5)
  8. LMArena Text Coding(style control なし)第三者・lmarena-ai/leaderboard-dataset text/latest-00000-of-00001.parquet (CC BY 4.0)計測・報告: LMArena(5)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. OTIS Mock AIME 2024-20255. SimpleQA Verified6. Chess Puzzles(Epoch AI)7. LMArena Text(style control・overall)8. LMArena Text Coding(style control なし)
1Gemini 4 CarbonHigh838/1496.5%90.5%95.1%100.0%77.7%78.0%1,5261,563
2GPT-6 AstraMax599/1495.8%93.7%97.6%100.0%75.6%74.0%1,4741,494
3Gemini 4 ArgonHigh559/1496.4%86.0%78.1%100.0%76.0%59.0%1,5221,540
4GPT-6.1 SolMax539/1495.4%93.7%100.0%100.0%73.9%68.0%1,4821,486
5GPT-6 LunaMax09/1490.5%79.0%56.1%98.9%41.4%36.0%1,4411,432
第三者・5 モデル・5 件 × 7 グループ(カバレッジ 7/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、OTIS Mock AIME 2024-2025、Chess Puzzles(Epoch AI)、MirrorCode、τ²-bench banking_knowledge (v1.0.1)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(5)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(5)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(5)
  4. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(5)
  5. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(5)
  6. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(5)
  7. τ²-bench banking_knowledge (v1.0.1)第三者・τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 alltools)計測・報告: Sierra(5)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. OTIS Mock AIME 2024-20255. Chess Puzzles(Epoch AI)6. MirrorCode7. τ²-bench banking_knowledge (v1.0.1)
1Claude Fable 5Max777/1485.9%87.0%90.2%99.7%47.0%70.0%39.7%
2Claude Opus 4.8Max708/1491.0%80.0%56.1%98.3%52.0%65.6%39.7%
3Qwen3.8-MaxXhigh618/1492.7%74.7%46.3%99.4%46.0%52.2%55.2%
4Kimi K3Max398/1493.1%72.2%39.0%97.2%36.0%54.4%37.1%
5Claude Opus 4.6Max48/1488.4%66.0%26.8%91.1%21.0%33.3%27.3%
第三者・5 モデル・5 件 × 6 グループ(カバレッジ 6/14 ベンチマーク)GPQA Diamond、OTIS Mock AIME 2024-2025、SimpleQA Verified、Chess Puzzles(Epoch AI)、MirrorCode、τ²-bench banking_knowledge (v1.0.1)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(5)
  2. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(5)
  3. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(5)
  4. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(5)
  5. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(5)
  6. τ²-bench banking_knowledge (v1.0.1)第三者・τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 alltools)計測・報告: Sierra(5)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. OTIS Mock AIME 2024-20253. SimpleQA Verified4. Chess Puzzles(Epoch AI)5. MirrorCode6. τ²-bench banking_knowledge (v1.0.1)
1Claude Opus 4.8Max758/1491.0%98.3%53.0%52.0%65.6%39.7%
2Qwen3.8-MaxXhigh588/1492.7%99.4%45.8%46.0%52.2%55.2%
3Kimi K3Max548/1493.1%97.2%50.6%36.0%54.4%37.1%
4Gemini 3.1 Pro PreviewHigh506/1494.4%95.6%73.5%51.0%8.9%26.0%
5Claude Opus 4.6Max138/1488.4%91.1%47.0%21.0%33.3%27.3%
第三者・5 モデル・5 件 × 6 グループ(カバレッジ 6/14 ベンチマーク)FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、SimpleQA Verified、Chess Puzzles(Epoch AI)、MirrorCode、τ²-bench banking_knowledge (v1.0.1)
  1. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(5)
  2. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(5)
  3. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(5)
  4. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(5)
  5. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(5)
  6. τ²-bench banking_knowledge (v1.0.1)第三者・τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 alltools)計測・報告: Sierra(5)
順位モデルeffort平均百分位値のあるベンチマーク1. FrontierMath Tiers 1-3(v2)2. FrontierMath Tier 4(v2)3. SimpleQA Verified4. Chess Puzzles(Epoch AI)5. MirrorCode6. τ²-bench banking_knowledge (v1.0.1)
1GPT-5.5Xhigh796/1485.3%72.5%63.0%63.0%23.3%44.6%
2Claude Opus 4.8Max758/1480.0%56.1%53.0%52.0%65.6%39.7%
3Qwen3.8-MaxXhigh508/1474.7%46.3%45.8%46.0%52.2%55.2%
4Kimi K3Max388/1472.2%39.0%50.6%36.0%54.4%37.1%
5Claude Opus 4.6Max88/1466.0%26.8%47.0%21.0%33.3%27.3%
第三者・5 モデル・5 件 × 3 グループ(カバレッジ 3/14 ベンチマーク)MirrorCode、LMArena Text(style control・overall)、LMArena Text Coding(style control なし)
  1. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(5)
  2. LMArena Text(style control・overall)第三者・lmarena-ai/leaderboard-dataset text_style_control/latest-00000-of-00001.parquet (CC BY 4.0)計測・報告: LMArena(5)
  3. LMArena Text Coding(style control なし)第三者・lmarena-ai/leaderboard-dataset text/latest-00000-of-00001.parquet (CC BY 4.0)計測・報告: LMArena(5)
順位モデルeffort平均百分位値のあるベンチマーク1. MirrorCode2. LMArena Text(style control・overall)3. LMArena Text Coding(style control なし)
1Claude Opus 5.5High923/1478.9%1,5061,542
2Gemini 4 ArgonHigh679/1458.9%1,5221,540
3GPT-6.1 SolMax509/1461.1%1,4821,486
4GPT-6 AstraMax429/1460.0%1,4741,494
5GPT-6 LunaMax09/1426.7%1,4411,432
第三者・4 モデル・4 件 × 9 グループ(カバレッジ 9/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、OTIS Mock AIME 2024-2025、SimpleQA Verified、Chess Puzzles(Epoch AI)、MirrorCode、LMArena Text(style control・overall)、LMArena Text Coding(style control なし)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(4)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(4)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(4)
  4. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(4)
  5. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(4)
  6. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(4)
  7. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(4)
  8. LMArena Text(style control・overall)第三者・lmarena-ai/leaderboard-dataset text_style_control/latest-00000-of-00001.parquet (CC BY 4.0)計測・報告: LMArena(4)
  9. LMArena Text Coding(style control なし)第三者・lmarena-ai/leaderboard-dataset text/latest-00000-of-00001.parquet (CC BY 4.0)計測・報告: LMArena(4)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. OTIS Mock AIME 2024-20255. SimpleQA Verified6. Chess Puzzles(Epoch AI)7. MirrorCode8. LMArena Text(style control・overall)9. LMArena Text Coding(style control なし)
1GPT-6 AstraMax699/1495.8%93.7%97.6%100.0%75.6%74.0%60.0%1,4741,494
2Gemini 4 ArgonHigh679/1496.4%86.0%78.1%100.0%76.0%59.0%58.9%1,5221,540
3GPT-6.1 SolMax659/1495.4%93.7%100.0%100.0%73.9%68.0%61.1%1,4821,486
4GPT-6 LunaMax09/1490.5%79.0%56.1%98.9%41.4%36.0%26.7%1,4411,432
第三者・4 モデル・4 件 × 8 グループ(カバレッジ 8/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、OTIS Mock AIME 2024-2025、SimpleQA Verified、Chess Puzzles(Epoch AI)、MirrorCode、τ²-bench banking_knowledge (v1.0.1)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(4)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(4)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(4)
  4. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(4)
  5. SimpleQA Verified第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0計測・報告: Epoch AI(4)
  6. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(4)
  7. MirrorCode第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0計測・報告: Epoch AI(4)
  8. τ²-bench banking_knowledge (v1.0.1)第三者・τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 alltools)計測・報告: Sierra(4)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. OTIS Mock AIME 2024-20255. SimpleQA Verified6. Chess Puzzles(Epoch AI)7. MirrorCode8. τ²-bench banking_knowledge (v1.0.1)
1Claude Opus 4.8Max838/1491.0%80.0%56.1%98.3%53.0%52.0%65.6%39.7%
2Qwen3.8-MaxXhigh638/1492.7%74.7%46.3%99.4%45.8%46.0%52.2%55.2%
3Kimi K3Max508/1493.1%72.2%39.0%97.2%50.6%36.0%54.4%37.1%
4Claude Opus 4.6Max48/1488.4%66.0%26.8%91.1%47.0%21.0%33.3%27.3%
第三者・4 モデル・4 件 × 5 グループ(カバレッジ 5/14 ベンチマーク)GPQA Diamond、FrontierMath Tiers 1-3(v2)、FrontierMath Tier 4(v2)、OTIS Mock AIME 2024-2025、Epoch Capabilities Index(ECI)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(4)
  2. FrontierMath Tiers 1-3(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0計測・報告: Epoch AI(4)
  3. FrontierMath Tier 4(v2)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0計測・報告: Epoch AI(4)
  4. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(4)
  5. Epoch Capabilities Index(ECI)第三者・Epoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)計測・報告: Epoch AI(4)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. FrontierMath Tiers 1-3(v2)3. FrontierMath Tier 4(v2)4. OTIS Mock AIME 2024-20255. Epoch Capabilities Index(ECI)
1DeepSeek-V4.1-Flash記載なし908/1489.8%67.4%26.8%98.3%156.7
2Gemini 3.1 Pro Preview記載なし775/1494.1%59.7%26.8%95.6%156.18
3Gemini 3 Flash Preview記載なし275/1483.2%51.2%17.1%92.8%152.32
4Gemini 2.5 Pro記載なし75/1485.3%24.6%0.0%84.7%144.87
第三者・4 モデル・4 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)GPQA Diamond、OTIS Mock AIME 2024-2025、Chess Puzzles(Epoch AI)、Epoch Capabilities Index(ECI)
  1. GPQA Diamond第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0計測・報告: Epoch AI(4)
  2. OTIS Mock AIME 2024-2025第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0計測・報告: Epoch AI(4)
  3. Chess Puzzles(Epoch AI)第三者・Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0計測・報告: Epoch AI(4)
  4. Epoch Capabilities Index(ECI)第三者・Epoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)計測・報告: Epoch AI(4)
順位モデルeffort平均百分位値のあるベンチマーク1. GPQA Diamond2. OTIS Mock AIME 2024-20253. Chess Puzzles(Epoch AI)4. Epoch Capabilities Index(ECI)
1DeepSeek-V4.1-Flash記載なし928/1489.8%98.3%48.0%156.7
2Claude Opus 4.8記載なし754/1485.4%84.4%19.0%159.94
3Claude Opus 4.5記載なし254/1480.7%48.1%5.0%151.92
4Claude Sonnet 4.5記載なし89/1473.7%35.6%6.0%147.63
第三者・3 モデル・3 件 × 4 グループ(カバレッジ 4/14 ベンチマーク)τ²-bench airline (v1.0.1)、τ²-bench retail (v1.0.1)、τ²-bench telecom (v1.0.1)、τ²-bench banking_knowledge (v1.0.1)
  1. τ²-bench airline (v1.0.1)第三者・τ²-bench v1.0.1(Sierra の実行)計測・報告: Sierra(3)
  2. τ²-bench retail (v1.0.1)第三者・τ²-bench v1.0.1(Sierra の実行)計測・報告: Sierra(3)
  3. τ²-bench telecom (v1.0.1)第三者・τ²-bench v1.0.1(Sierra の実行)計測・報告: Sierra(3)
  4. τ²-bench banking_knowledge (v1.0.1)第三者・τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 terminal)計測・報告: Sierra(3)
順位モデルeffort平均百分位値のあるベンチマーク1. τ²-bench airline (v1.0.1)2. τ²-bench retail (v1.0.1)3. τ²-bench telecom (v1.0.1)4. τ²-bench banking_knowledge (v1.0.1)
1Gemini 3 Flash PreviewHigh1008/1482.5%76.8%91.2%27.3%
2Gemini 3 Pro PreviewHigh384/1480.5%75.9%91.0%18.0%
3Claude Sonnet 4.5記載なし139/1472.0%72.4%84.9%25.3%

コストバランス

出典のコストはベンチマークごとに意味が違うため、ベンチマークをまたいで合算せず、比較グループの中だけで見ます。

  • 前線: 同じかそれより安いコストで、より良いスコアの点が無い点(パレート前線)。安い順に並べます。出典のコストの無い点は判定から外します。
  • effort の差分: 同じモデルで値のある隣どうしの effort の間の、スコアの差(% の指標はポイント)とコストの倍率。間の effort に値が無いときは、その旨を添えます。両端で計測・報告が違うときは「前 → 後」で示します。

Terminal-Bench 4.0

公式の結果(Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ))

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
Claude Sonnet 5.5Low20.0%$0.76Cost per attemptAnthropic
Claude Sonnet 5.5Med28.8%$0.83Cost per attemptAnthropic
Claude Opus 5.5Low38.5%$1.29Cost per attemptAnthropic
Claude Sonnet 5.5High43.0%$1.94Cost per attemptAnthropic
Claude Opus 5.5Med57.6%$2.94Cost per attemptAnthropic
Claude Opus 5.5High64.2%$3.88Cost per attemptAnthropic
Claude Opus 5.5Xhigh66.4%$7.35Cost per attemptAnthropic
Claude Sonnet 5.5Max70.6%$12.54Cost per attemptAnthropic
effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Opus 5.5Low → Med+19.1pt×2.28Anthropic
Claude Opus 5.5Med → High+6.6pt×1.32Anthropic
Claude Opus 5.5High → Xhigh+2.2pt×1.89Anthropic
Claude Opus 5.5Xhigh → Max-1.6pt×1.53Anthropic
Claude Sonnet 5.5Low → Med+8.8pt×1.09Anthropic
Claude Sonnet 5.5Med → High+14.2pt×2.34Anthropic
Claude Sonnet 5.5High → Xhigh+18.5pt×2.73Anthropic
Claude Sonnet 5.5Xhigh → Max+9.1pt×2.37Anthropic
Claude Fable 5.1Low → Med+3.2pt×1.37Anthropic
Claude Fable 5.1Med → High+6.0pt×1.35Anthropic
Claude Fable 5.1High → Xhigh+1.9pt×1.5Anthropic
Claude Fable 5.1Xhigh → Max+4.5pt×1.23Anthropic
Claude Opus 5Low → Med+12.7pt×1.65Anthropic
Claude Opus 5Med → High+5.8pt×1.52Anthropic
Claude Opus 5High → Xhigh+3.6pt×1.27Anthropic
Claude Opus 5Xhigh → Max+1.7pt×1.17Anthropic
Claude Sonnet 5Low → Med+1.2pt×1.28Anthropic
Claude Sonnet 5Med → High+0.1pt×1.7Anthropic
Claude Sonnet 5High → Xhigh+2.5pt×1.21Anthropic
Claude Sonnet 5Xhigh → Max+3.3pt×1.17Anthropic
GPT-6 AstraLow → Med+4.2pt×1.24OpenAI
GPT-6 AstraMed → High+4.0pt×1.17OpenAI
GPT-6 AstraHigh → Xhigh-0.3pt×1.04OpenAI
GPT-6 AstraXhigh → Max-0.9pt×1.38OpenAI
GPT-5.6 SolLow → Med+13.0pt×1.84OpenAI
GPT-5.6 SolMed → High+5.2pt×1.53OpenAI
GPT-5.6 SolHigh → Xhigh+2.4pt×1.31OpenAI
GPT-5.6 SolXhigh → Max+8.8pt×1.46OpenAI

公式の結果(Introducing Claude Haiku 5.5)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
Claude Haiku 5.5Low12.7%$0.42Cost per attemptAnthropic
Claude Sonnet 5.5Low20.0%$0.62Cost per attemptAnthropic
Claude Sonnet 5.5Med28.8%$0.68Cost per attemptAnthropic
Claude Sonnet 5.5High43.0%$1.46Cost per attemptAnthropic
Claude Sonnet 5.5Xhigh61.5%$4.34Cost per attemptAnthropic
Claude Sonnet 5.5Max70.6%$10.44Cost per attemptAnthropic

出典のコストの無い 1 点は前線の判定から外しています。

effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Sonnet 5.5Low → Med+8.8pt×1.1Anthropic
Claude Sonnet 5.5Med → High+14.2pt×2.15Anthropic
Claude Sonnet 5.5High → Xhigh+18.5pt×2.97Anthropic
Claude Sonnet 5.5Xhigh → Max+9.1pt×2.41Anthropic
Claude Haiku 5.5Low → Med+7.6pt×1.62Anthropic
Claude Haiku 5.5Med → High+4.5pt×1.53Anthropic
Claude Haiku 5.5High → Xhigh+6.7pt×1.68Anthropic
Claude Haiku 5.5Xhigh → Max+7.7pt×1.51Anthropic

FrontierCode v1.1(main set)

公式の結果(Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ))

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
Claude Sonnet 5.5Low29.3%$0.19Cost per taskAnthropic
Claude Sonnet 5.5Med36.5%$0.24Cost per taskAnthropic
Claude Opus 5.5Low47.3%$0.40Cost per taskAnthropic
Claude Sonnet 5.5High49.4%$0.42Cost per taskAnthropic
Claude Opus 5.5Med54.6%$0.80Cost per taskAnthropic
effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Opus 5.5Low → Med+7.3pt×2.0Anthropic
Claude Opus 5.5Med → High-0.6pt×1.36Anthropic
Claude Opus 5.5High → Xhigh-2.6pt×2.06Anthropic
Claude Opus 5.5Xhigh → Max+3.0pt×2.75Anthropic
Claude Sonnet 5.5Low → Med+7.2pt×1.26Anthropic
Claude Sonnet 5.5Med → High+12.9pt×1.75Anthropic
Claude Sonnet 5.5High → Xhigh+2.7pt×3.79Anthropic
Claude Sonnet 5.5Xhigh → Max-5.9pt×13.07Anthropic
Claude Fable 5.1Low → Med-1.9pt×1.33Anthropic
Claude Fable 5.1Med → High-0.6pt×1.61Anthropic
Claude Fable 5.1High → Xhigh-1.6pt×1.76Anthropic
Claude Fable 5.1Xhigh → Max+1.6pt×1.38Anthropic
Claude Opus 5Low → Med+11.4pt×1.75Anthropic
Claude Opus 5Med → High-5.4pt×1.65Anthropic
Claude Opus 5High → Xhigh-4.4pt×1.18Anthropic
Claude Opus 5Xhigh → Max+4.4pt×1.37Anthropic
Claude Sonnet 5Low → Med+6.5pt×1.59Anthropic
Claude Sonnet 5Med → High+4.2pt×1.6Anthropic
Claude Sonnet 5High → Xhigh+3.3pt×1.65Anthropic
Claude Sonnet 5Xhigh → Max-0.3pt×1.7Anthropic
GPT-6 AstraLow → Med+3.5pt×1.43記載なし
GPT-6 AstraMed → High+2.1pt×1.25記載なし
GPT-6 AstraHigh → Xhigh-0.3pt×1.09記載なし
GPT-6 AstraXhigh → Max+2.7pt×1.41記載なし
GPT-6 SolLow → Med+8.6pt×1.79記載なし
GPT-6 SolMed → High+1.8pt×1.35記載なし
GPT-6 SolHigh → Xhigh+0.7pt×1.27記載なし
GPT-6 SolXhigh → Max+0.9pt×1.57記載なし
GPT-5.6 SolLow → Med+4.5pt×1.43記載なし
GPT-5.6 SolMed → High+5.2pt×1.3記載なし
GPT-5.6 SolHigh → Xhigh+1.7pt×1.19記載なし
GPT-5.6 SolXhigh → Max+0.7pt×1.25記載なし

FrontierCode 2.0

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
Claude Opus 5.5Low14.7%$0.75Cost per taskAnthropic
Claude Opus 5.5Med30.1%$1.70Cost per taskAnthropic
Claude Fable 5.2Low34.2%$2.10Cost per taskAnthropic
Claude Opus 5.5High34.5%$2.42Cost per taskAnthropic
Claude Fable 5.2Med37.6%$2.48Cost per taskAnthropic
GPT-6 AstraHigh38.3%$4.28Cost per task記載なし
GPT-6 AstraXhigh41.4%$4.54Cost per task記載なし
Claude Opus 5.5Xhigh43.4%$5.18Cost per taskAnthropic
Claude Fable 5.2High45.3%$6.72Cost per taskAnthropic
Claude Opus 5.5Max45.5%$12.82Cost per taskAnthropic
Claude Fable 5.2Xhigh49.1%$14.19Cost per taskAnthropic
effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Fable 5.2Low → Med+3.4pt×1.18Anthropic
Claude Fable 5.2Med → High+7.7pt×2.71Anthropic
Claude Fable 5.2High → Xhigh+3.8pt×2.11Anthropic
Claude Fable 5.2Xhigh → Max-3.5pt×3.15Anthropic
Claude Opus 5.5Low → Med+15.4pt×2.27Anthropic
Claude Opus 5.5Med → High+4.4pt×1.42Anthropic
Claude Opus 5.5High → Xhigh+8.9pt×2.14Anthropic
Claude Opus 5.5Xhigh → Max+2.1pt×2.47Anthropic
Claude Fable 5.1Low → Med+0.9pt×1.29Anthropic
Claude Fable 5.1Med → High+4.2pt×1.39Anthropic
Claude Fable 5.1High → Xhigh+6.7pt×1.61Anthropic
Claude Fable 5.1Xhigh → Max-0.7pt×1.32Anthropic
GPT-6 AstraLow → Med+5.7pt×1.43記載なし
GPT-6 AstraMed → High+1.5pt×1.23記載なし
GPT-6 AstraHigh → Xhigh+3.1pt×1.06記載なし
GPT-6 AstraXhigh → Max+1.4pt×1.47記載なし
Gemini 4 CarbonLow → Med+8.7pt×0.94記載なし
Gemini 4 CarbonMed → High-2.7pt×1.13記載なし

CursorBench 4.0

公式の結果(Introducing Claude Opus 5.5)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
GPT-5.6 SolLow24.6%$0.87Cost per task記載なし
Claude Opus 5.5Low43.7%$1.18Cost per taskAnthropic
Claude Opus 5.5Med52.5%$2.90Cost per taskAnthropic
Claude Opus 5.5High56.0%$3.97Cost per taskAnthropic
Claude Opus 5.5Xhigh56.0%$6.99Cost per taskAnthropic
Claude Opus 5.5Max57.8%$13.43Cost per taskAnthropic
effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Opus 5.5Low → Med+8.8pt×2.46Anthropic
Claude Opus 5.5Med → High+3.5pt×1.37Anthropic
Claude Opus 5.5High → Xhigh±0.0pt×1.76Anthropic
Claude Opus 5.5Xhigh → Max+1.8pt×1.92Anthropic
Claude Fable 5.1Low → Med+1.7pt×1.3Anthropic
Claude Fable 5.1Med → High+2.4pt×1.29Anthropic
Claude Fable 5.1High → Xhigh+2.4pt×1.43Anthropic
Claude Fable 5.1Xhigh → Max+0.2pt×1.33Anthropic
Claude Opus 5Low → Med+2.6pt×1.43Anthropic
Claude Opus 5Med → High+1.4pt×1.3Anthropic
Claude Opus 5High → Xhigh+1.4pt×1.27Anthropic
Claude Opus 5Xhigh → Max+0.5pt×1.05Anthropic
GPT-5.6 SolLow → Med+6.5pt×2.03記載なし
GPT-5.6 SolMed → High+4.6pt×1.61記載なし
GPT-5.6 SolHigh → Xhigh+2.0pt×1.54記載なし
GPT-5.6 SolXhigh → Max+4.0pt×1.87記載なし

公式の結果(Introducing Claude Sonnet 5.5)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
Claude Sonnet 5.5Low35.8%$0.50Cost per taskAnthropic
Claude Sonnet 5.5Med39.2%$0.70Cost per taskAnthropic
Claude Opus 5.5Low43.7%$1.17Cost per taskAnthropic
Claude Sonnet 5.5High47.8%$1.67Cost per taskAnthropic
Claude Opus 5.5Med52.5%$2.91Cost per taskAnthropic
Claude Sonnet 5.5Xhigh53.1%$3.88Cost per taskAnthropic
Claude Opus 5.5High56.0%$3.97Cost per taskAnthropic
Claude Opus 5.5Xhigh56.0%$6.98Cost per taskAnthropic
Claude Opus 5.5Max57.8%$13.43Cost per taskAnthropic
effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Opus 5.5Low → Med+8.8pt×2.49Anthropic
Claude Opus 5.5Med → High+3.5pt×1.36Anthropic
Claude Opus 5.5High → Xhigh±0.0pt×1.76Anthropic
Claude Opus 5.5Xhigh → Max+1.8pt×1.92Anthropic
Claude Sonnet 5.5Low → Med+3.4pt×1.4Anthropic
Claude Sonnet 5.5Med → High+8.6pt×2.39Anthropic
Claude Sonnet 5.5High → Xhigh+5.3pt×2.32Anthropic
Claude Sonnet 5.5Xhigh → Max+2.4pt×2.49Anthropic
Claude Sonnet 5Low → Med+3.9pt×1.66Anthropic
Claude Sonnet 5Med → High+2.8pt×1.51Anthropic
Claude Sonnet 5High → Xhigh+1.2pt×1.31Anthropic
Claude Sonnet 5Xhigh → Max+2.1pt×1.58Anthropic
GPT-5.6 SolLow → Med+6.5pt×2.03記載なし
GPT-5.6 SolMed → High+4.6pt×1.61記載なし
GPT-5.6 SolHigh → Xhigh+2.0pt×1.54記載なし
GPT-5.6 SolXhigh → Max+4.0pt×1.87記載なし

GDPval-AA v2.1

公式の結果(Introducing Claude Opus 5.5)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
Claude Opus 5.5Low1,224$0.21Estimated cost per taskAnthropic
GPT-5.6 SolLow1,289$0.27Estimated cost per task記載なし
Claude Opus 5Low1,294$0.58Estimated cost per taskAnthropic
GPT-5.6 SolMed1,403$0.60Estimated cost per task記載なし
Claude Opus 5.5Med1,576$0.86Estimated cost per taskAnthropic
Claude Opus 5.5High1,692$1.54Estimated cost per taskAnthropic
Claude Opus 5.5Xhigh1,820$4.21Estimated cost per taskAnthropic
Claude Opus 5.5Max1,846$8.92Estimated cost per taskAnthropic
effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Opus 5.5Low → Med+352×4.1Anthropic
Claude Opus 5.5Med → High+116×1.79Anthropic
Claude Opus 5.5High → Xhigh+128×2.73Anthropic
Claude Opus 5.5Xhigh → Max+26×2.12Anthropic
Claude Fable 5.1Low → Med+86×1.54Anthropic
Claude Fable 5.1Med → High+81×1.58Anthropic
Claude Fable 5.1High → Xhigh+104×2.07Anthropic
Claude Fable 5.1Xhigh → Max+14×1.35Anthropic
Claude Opus 5Low → Med+182×2.34Anthropic
Claude Opus 5Med → High+105×2.23Anthropic
Claude Opus 5High → Xhigh+95×1.64Anthropic
Claude Opus 5Xhigh → Max+32×1.36Anthropic
GPT-6 AstraLow → Med+102×2.14記載なし
GPT-6 AstraMed → High+17×1.34記載なし
GPT-6 AstraHigh → Xhigh+31×1.25記載なし
GPT-6 AstraXhigh → Max+26×1.49記載なし
GPT-5.6 SolLow → Med+114×2.22記載なし
GPT-5.6 SolMed → High+77×1.85記載なし
GPT-5.6 SolHigh → Xhigh+68×1.53記載なし
GPT-5.6 SolXhigh → Max+40×1.65記載なし

公式の結果(Introducing Claude Haiku 5.5)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
GPT-6 LunaLow1,036$0.004Cost per task記載なし
Claude Haiku 5.5Low1,125$0.01Cost per taskAnthropic
GPT-6 LunaMed1,262$0.02Cost per task記載なし
GPT-6 LunaHigh1,344$0.03Cost per task記載なし
GPT-6 LunaXhigh1,364$0.05Cost per task記載なし
GPT-6 LunaMax1,437$0.09Cost per task記載なし
Claude Haiku 5.5Xhigh1,513$0.27Cost per taskAnthropic
Claude Sonnet 5.5High1,551$0.62Cost per taskAnthropic
Claude Haiku 5.5Max1,620$0.87Cost per taskAnthropic
Claude Sonnet 5.5Xhigh1,731$1.88Cost per taskAnthropic
Claude Sonnet 5.5Max1,840$6.78Cost per taskAnthropic
effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Sonnet 5.5Low → Med+145×1.23Anthropic
Claude Sonnet 5.5Med → High+227×2.3Anthropic
Claude Sonnet 5.5High → Xhigh+180×3.03Anthropic
Claude Sonnet 5.5Xhigh → Max+109×3.61Anthropic
Claude Haiku 5.5Low → Med+152×3.0Anthropic
Claude Haiku 5.5Med → High+143×3.0Anthropic
Claude Haiku 5.5High → Xhigh+93×3.0Anthropic
Claude Haiku 5.5Xhigh → Max+107×3.22Anthropic
GPT-6 LunaLow → Med+226×5.0記載なし
GPT-6 LunaMed → High+82×1.5記載なし
GPT-6 LunaHigh → Xhigh+20×1.67記載なし
GPT-6 LunaXhigh → Max+73×1.8記載なし

AA-Briefcase v1.1(Elo)

公式の結果(Introducing Claude Sonnet 5.5)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
GPT-6 SolLow905$0.12Cost per task記載なし
GPT-6 SolMed1,142$0.34Cost per task記載なし
GPT-6 SolHigh1,289$0.63Cost per task記載なし
GPT-6 SolXhigh1,364$1.19Cost per task記載なし
Claude Sonnet 5.5Med1,461$1.64Cost per taskAnthropic
GPT-6 SolMax1,483$2.67Cost per task記載なし
Claude Sonnet 5.5High1,634$3.95Cost per taskAnthropic
Claude Opus 5.5Med1,642$4.40Cost per taskAnthropic
Claude Opus 5.5High1,705$6.27Cost per taskAnthropic
Claude Sonnet 5.5Xhigh1,746$9.63Cost per taskAnthropic
Claude Opus 5.5Xhigh1,780$12.27Cost per taskAnthropic
Claude Opus 5.5Max1,822$21.05Cost per taskAnthropic
effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Opus 5.5Low → Med+357×3.83Anthropic
Claude Opus 5.5Med → High+63×1.42Anthropic
Claude Opus 5.5High → Xhigh+75×1.96Anthropic
Claude Opus 5.5Xhigh → Max+42×1.72Anthropic
Claude Sonnet 5.5Low → Med+197×1.89Anthropic
Claude Sonnet 5.5Med → High+173×2.41Anthropic
Claude Sonnet 5.5High → Xhigh+112×2.44Anthropic
Claude Sonnet 5.5Xhigh → Max+65×3.03Anthropic
Claude Sonnet 5Low → Med+133×2.11Anthropic
Claude Sonnet 5Med → High+121×2.17Anthropic
Claude Sonnet 5High → Xhigh+97×2.01Anthropic
Claude Sonnet 5Xhigh → Max+85×1.91Anthropic
GPT-6 SolLow → Med+237×2.83記載なし
GPT-6 SolMed → High+147×1.85記載なし
GPT-6 SolHigh → Xhigh+75×1.89記載なし
GPT-6 SolXhigh → Max+119×2.24記載なし

AutomationBench

公式の結果(Introducing Claude Opus 5.5)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
GPT-5.6 SolLow11.7%$0.42Cost per taskZapier
Claude Opus 5.5Low23.3%$0.50Cost per taskZapier
Claude Opus 5.5Med28.6%$0.64Cost per taskZapier
Claude Opus 5.5High32.0%$0.70Cost per taskZapier
Claude Opus 5.5Xhigh34.4%$0.86Cost per taskZapier
Claude Opus 5.5Max40.0%$1.37Cost per taskZapier
GPT-6 AstraMax41.4%$1.77Cost per taskZapier

出典のコストの無い 1 点は前線の判定から外しています。

effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Opus 5.5Low → Med+5.3pt×1.28Zapier
Claude Opus 5.5Med → High+3.4pt×1.09Zapier
Claude Opus 5.5High → Xhigh+2.4pt×1.23Zapier
Claude Opus 5.5Xhigh → Max+5.6pt×1.59Zapier
Claude Opus 5Low → Med+3.5pt×1.19Zapier
Claude Opus 5Med → High-3.3pt×1.16Zapier
Claude Opus 5High → Xhigh+4.7pt×1.12Zapier
Claude Opus 5Xhigh → Max+1.6pt×1.1Zapier
GPT-6 AstraLow → Med+3.8pt×1.19Zapier
GPT-6 AstraMed → High+3.0pt×1.13Zapier
GPT-6 AstraHigh → Xhigh+1.9pt×1.06Zapier
GPT-6 AstraXhigh → Max+2.4pt×1.16Zapier
GPT-5.6 SolLow → Med+7.9pt×1.36Zapier
GPT-5.6 SolMed → High+5.2pt×1.12Zapier
GPT-5.6 SolHigh → Xhigh+1.5pt×1.16Zapier
GPT-5.6 SolXhigh → Max+2.5pt×1.23Zapier

Humanity's Last Exam(ツールあり)

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
Claude Opus 5.5Low57.4%$0.08Cost per attemptAnthropic
Claude Opus 5.5Med63.0%$0.17Cost per attemptAnthropic
Claude Fable 5.2Low68.4%$0.19Cost per attemptAnthropic
Claude Fable 5.2Med68.6%$0.21Cost per attemptAnthropic
Claude Fable 5.2High71.5%$0.63Cost per attemptAnthropic
Claude Fable 5.2Xhigh73.4%$1.48Cost per attemptAnthropic
Claude Fable 5.2Max74.2%$4.51Cost per attemptAnthropic

出典のコストの無い 1 点は前線の判定から外しています。

effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Fable 5.2Low → Med+0.2pt×1.11Anthropic
Claude Fable 5.2Med → High+2.9pt×3.0Anthropic
Claude Fable 5.2High → Xhigh+1.9pt×2.35Anthropic
Claude Fable 5.2Xhigh → Max+0.8pt×3.05Anthropic
Claude Opus 5.5Low → Med+5.6pt×2.13Anthropic
Claude Opus 5.5Med → High+0.9pt×1.71Anthropic
Claude Opus 5.5High → Xhigh+2.5pt×2.38Anthropic
Claude Opus 5.5Xhigh → Max+1.3pt×3.16Anthropic
Claude Fable 5.1Low → Med+2.2pt×1.44Anthropic
Claude Fable 5.1Med → High+0.7pt×1.61Anthropic
Claude Fable 5.1High → Xhigh+2.1pt×1.97Anthropic
Claude Fable 5.1Xhigh → Max+0.4pt×1.4Anthropic

Humanity's Last Exam(ツールなし)

公式の結果(Introducing Claude Haiku 5.5)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
Claude Haiku 5.5Low30.9%$0.003Cost per attemptAnthropic
Claude Haiku 5.5Med35.5%$0.007Cost per attemptAnthropic
Claude Haiku 5.5High39.7%$0.02Cost per attemptAnthropic
Claude Sonnet 5.5Med43.2%$0.03Cost per attemptAnthropic
Claude Haiku 5.5Xhigh44.1%$0.06Cost per attemptAnthropic
Claude Sonnet 5.5High47.9%$0.08Cost per attemptAnthropic
Claude Sonnet 5.5Xhigh53.0%$0.20Cost per attemptAnthropic
Claude Sonnet 5.5Max56.9%$0.89Cost per attemptAnthropic
effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Sonnet 5.5Low → Med+1.8pt×1.0Anthropic
Claude Sonnet 5.5Med → High+4.7pt×2.67Anthropic
Claude Sonnet 5.5High → Xhigh+5.1pt×2.5Anthropic
Claude Sonnet 5.5Xhigh → Max+3.9pt×4.45Anthropic
Claude Haiku 5.5Low → Med+4.6pt×2.33Anthropic
Claude Haiku 5.5Med → High+4.2pt×2.86Anthropic
Claude Haiku 5.5High → Xhigh+4.4pt×3.0Anthropic
Claude Haiku 5.5Xhigh → Max+1.8pt×3.5Anthropic

OSWorld 2.1(offline subset)

公式の結果(Introducing Claude Haiku 5.5)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
GPT-6 LunaLow19.2%$0.04Cost per attempt記載なし
Claude Haiku 5.5Low42.0%$0.07Cost per attemptAnthropic
Claude Haiku 5.5Med53.3%$0.13Cost per attemptAnthropic
Claude Haiku 5.5High61.3%$0.18Cost per attemptAnthropic
Claude Haiku 5.5Xhigh67.6%$0.28Cost per attemptAnthropic
Claude Haiku 5.5Max72.4%$0.61Cost per attemptAnthropic
Claude Sonnet 5.5High73.2%$1.38Cost per attemptAnthropic
Claude Sonnet 5.5Xhigh81.1%$2.22Cost per attemptAnthropic
Claude Sonnet 5.5Max83.9%$5.73Cost per attemptAnthropic
effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Sonnet 5.5Low → Med+8.1pt×1.37Anthropic
Claude Sonnet 5.5Med → High+7.2pt×1.48Anthropic
Claude Sonnet 5.5High → Xhigh+7.9pt×1.61Anthropic
Claude Sonnet 5.5Xhigh → Max+2.8pt×2.58Anthropic
Claude Haiku 5.5Low → Med+11.3pt×1.86Anthropic
Claude Haiku 5.5Med → High+8.0pt×1.38Anthropic
Claude Haiku 5.5High → Xhigh+6.3pt×1.56Anthropic
Claude Haiku 5.5Xhigh → Max+4.8pt×2.18Anthropic
GPT-6 LunaLow → Med+18.3pt×3.25記載なし
GPT-6 LunaMed → High+4.8pt×1.08記載なし
GPT-6 LunaHigh → Xhigh+2.5pt×1.21記載なし
GPT-6 LunaXhigh → Max+4.1pt×1.18記載なし

Chartography(ツールなし)

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
Claude Opus 5.5Low57.0%$0.02Cost per attemptAnthropic
Claude Fable 5.2Med66.4%$0.03Cost per attemptAnthropic
Claude Fable 5.2High69.8%$0.07Cost per attemptAnthropic
Claude Fable 5.2Xhigh71.4%$0.15Cost per attemptAnthropic
Claude Fable 5.2Max73.4%$0.75Cost per attemptAnthropic

出典のコストの無い 2 点は前線の判定から外しています。

effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Fable 5.2Low → Med+2.2pt×1.0Anthropic
Claude Fable 5.2Med → High+3.4pt×2.33Anthropic
Claude Fable 5.2High → Xhigh+1.6pt×2.14Anthropic
Claude Fable 5.2Xhigh → Max+2.0pt×5.0Anthropic
Claude Opus 5.5Low → Med+2.0pt×1.5Anthropic
Claude Opus 5.5Med → High+0.2pt×1.0Anthropic
Claude Opus 5.5High → Xhigh+3.2pt×2.67Anthropic
Claude Opus 5.5Xhigh → Max+2.0pt×6.13Anthropic
Claude Fable 5.1Low → Med+5.6pt×1.67Anthropic
Claude Fable 5.1Med → High+3.8pt×1.6Anthropic
Claude Fable 5.1High → Xhigh+1.0pt×4.75Anthropic
Claude Fable 5.1Xhigh → Max-2.2pt×2.29Anthropic

WANDR

公式の結果(Introducing Claude Opus 5.5)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
Claude Opus 5.5Low31.2%$1.20Cost per attemptAnthropic
Claude Opus 5Low50.5%$10.65Cost per attemptAnthropic
Claude Opus 5.5Med62.8%$11.20Cost per attemptAnthropic
Claude Opus 5.5High67.3%$16.88Cost per attemptAnthropic
Claude Opus 5.5Xhigh71.3%$29.06Cost per attemptAnthropic
Claude Opus 5.5Max72.3%$37.92Cost per attemptAnthropic
effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Opus 5.5Low → Med+31.6pt×9.33Anthropic
Claude Opus 5.5Med → High+4.5pt×1.51Anthropic
Claude Opus 5.5High → Xhigh+4.0pt×1.72Anthropic
Claude Opus 5.5Xhigh → Max+1.0pt×1.3Anthropic
Claude Fable 5.1Low → Med+1.2pt×1.18Anthropic
Claude Fable 5.1Med → High+2.2pt×1.23Anthropic
Claude Fable 5.1High → Xhigh+1.0pt×1.27Anthropic
Claude Fable 5.1Xhigh → Max+1.0pt×1.15Anthropic
Claude Opus 5Low → Med+7.6pt×2.25Anthropic
Claude Opus 5Med → High+6.5pt×1.81Anthropic
Claude Opus 5High → Xhigh+2.4pt×1.24Anthropic
Claude Opus 5Xhigh → Max+0.2pt×1.14Anthropic

WANDR(relaxed score)

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
Claude Opus 5.5Low31.2%$1.25Cost per attemptAnthropic
Claude Opus 5.5Med62.8%$11.40Cost per attemptAnthropic
Claude Fable 5.2Low71.4%$12.66Cost per attemptAnthropic
Claude Fable 5.2Med73.6%$15.71Cost per attemptAnthropic
Claude Fable 5.2High79.3%$29.20Cost per attemptAnthropic
Claude Fable 5.2Xhigh80.6%$39.84Cost per attemptAnthropic
Claude Fable 5.2Max82.1%$53.11Cost per attemptAnthropic
effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Fable 5.2Low → Med+2.2pt×1.24Anthropic
Claude Fable 5.2Med → High+5.7pt×1.86Anthropic
Claude Fable 5.2High → Xhigh+1.3pt×1.36Anthropic
Claude Fable 5.2Xhigh → Max+1.5pt×1.33Anthropic
Claude Opus 5.5Low → Med+31.6pt×9.12Anthropic
Claude Opus 5.5Med → High+4.5pt×1.52Anthropic
Claude Opus 5.5High → Xhigh+4.0pt×1.69Anthropic
Claude Opus 5.5Xhigh → Max+1.0pt×1.32Anthropic
Claude Fable 5.1Low → Med+1.2pt×1.18Anthropic
Claude Fable 5.1Med → High+2.2pt×1.31Anthropic
Claude Fable 5.1High → Xhigh+1.0pt×1.18Anthropic
Claude Fable 5.1Xhigh → Max+1.0pt×1.15Anthropic

BenchCAD 2.1(ツールあり)

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
Claude Opus 5.5Low6.3$0.03Cost per attemptAnthropic
Claude Opus 5.5Med19.1$1.11Cost per attemptAnthropic
Claude Fable 5.2Low26.9$1.89Cost per attemptAnthropic
Claude Opus 5.5High29.7$2.57Cost per attemptAnthropic
Claude Fable 5.2Med32.4$2.66Cost per attemptAnthropic
Claude Opus 5.5Xhigh54.5$8.91Cost per attemptAnthropic
Claude Fable 5.2High61$9.99Cost per attemptAnthropic
Claude Opus 5.5Max65.3$19.86Cost per attemptAnthropic
Claude Fable 5.2Xhigh74.1$23.23Cost per attemptAnthropic
Claude Fable 5.2Max83.5$50.77Cost per attemptAnthropic
effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Fable 5.2Low → Med+5.5×1.41Anthropic
Claude Fable 5.2Med → High+28.6×3.76Anthropic
Claude Fable 5.2High → Xhigh+13.1×2.33Anthropic
Claude Fable 5.2Xhigh → Max+9.4×2.19Anthropic
Claude Opus 5.5Low → Med+12.8×37.0Anthropic
Claude Opus 5.5Med → High+10.6×2.32Anthropic
Claude Opus 5.5High → Xhigh+24.8×3.47Anthropic
Claude Opus 5.5Xhigh → Max+10.8×2.23Anthropic
Claude Fable 5.1Low → Med+3.6×1.93Anthropic
Claude Fable 5.1Med → High+10.1×2.18Anthropic
Claude Fable 5.1High → Xhigh+14.7×2.2Anthropic
Claude Fable 5.1Xhigh → Max+3.9×1.27Anthropic

CVE Rediscovery

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
Claude Opus 5.5Low59.4%$0.11Cost per attemptAnthropic
Claude Opus 5.5Med66.1%$0.28Cost per attemptAnthropic
Claude Mythos 5.2Low72.4%$0.34Cost per attemptAnthropic
Claude Mythos 5.2Med73.3%$0.40Cost per attemptAnthropic
Claude Mythos 5.2High83.2%$1.01Cost per attemptAnthropic
Claude Mythos 5.2Xhigh88.1%$2.40Cost per attemptAnthropic
Claude Mythos 5.2Max89.6%$9.13Cost per attemptAnthropic
effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Mythos 5.2Low → Med+0.9pt×1.18Anthropic
Claude Mythos 5.2Med → High+9.9pt×2.53Anthropic
Claude Mythos 5.2High → Xhigh+4.9pt×2.38Anthropic
Claude Mythos 5.2Xhigh → Max+1.5pt×3.8Anthropic
Claude Opus 5.5Low → Med+6.7pt×2.55Anthropic
Claude Opus 5.5Med → High+3.9pt×1.54Anthropic
Claude Opus 5.5High → Xhigh+6.4pt×2.44Anthropic
Claude Opus 5.5Xhigh → Max+4.9pt×2.84Anthropic
Claude Mythos 5.1Low → Med+3.9pt×1.51Anthropic
Claude Mythos 5.1Med → High+4.0pt×1.56Anthropic
Claude Mythos 5.1High → Xhigh+4.0pt×2.02Anthropic
Claude Mythos 5.1Xhigh → Max+0.5pt×1.47Anthropic

GPQA Diamond

第三者の結果(Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0)

出典のコストのある点が無いため、前線は出せません。

effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Fable 5Low → HighMed の値なし+4.5pt—Epoch AI
Claude Fable 5High → MaxXhigh の値なし+2.5pt—Epoch AI
Claude Opus 4.8Low → MaxMed・High・Xhigh の値なし+2.7pt—Epoch AI
Claude Opus 4.7Xhigh → Max-3.8pt—Epoch AI
Claude Sonnet 4.6Med → High±0.0pt—Epoch AI
Claude Sonnet 4.6High → Max-4.5pt—Epoch AI
GPT-5.6 SolNone → Low+7.1pt—Epoch AI
GPT-5.6 SolLow → MaxMed・High・Xhigh の値なし+3.6pt—Epoch AI
GPT-5.6 TerraNone → Low+10.1pt—Epoch AI
GPT-5.6 TerraLow → MaxMed・High・Xhigh の値なし+5.9pt—Epoch AI
GPT-5.6 LunaNone → Low+18.7pt—Epoch AI
GPT-5.6 LunaLow → MaxMed・High・Xhigh の値なし+9.3pt—Epoch AI
GPT-5.5None → Low+13.4pt—Epoch AI
Gemini 3.6 FlashMinimal → Low+0.5pt—Epoch AI
Gemini 3.6 FlashLow → HighMed の値なし+7.8pt—Epoch AI
Gemini 3.5 Flash-LiteMinimal → Low+1.5pt—Epoch AI
Gemini 3.5 Flash-LiteLow → HighMed の値なし+7.6pt—Epoch AI
Gemini 3.1 Flash-LiteMinimal → Low+0.5pt—Epoch AI
Gemini 3.1 Flash-LiteLow → HighMed の値なし+7.6pt—Epoch AI
Grok 4.6High → Xhigh-0.8pt—Epoch AI
Kimi K3Low → High+7.1pt—Epoch AI
Kimi K3High → Max+1.2pt—Epoch AI

FrontierMath Tiers 1-3(v2)

第三者の結果(Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0)

出典のコストのある点が無いため、前線は出せません。

effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
GPT-5.6 LunaNone → Low+1.8pt—Epoch AI
GPT-5.6 LunaLow → MaxMed・High・Xhigh の値なし+40.7pt—Epoch AI
Gemini 3.1 Flash-LiteMinimal → Low+1.1pt—Epoch AI
Gemini 3.1 Flash-LiteLow → HighMed の値なし+5.3pt—Epoch AI

FrontierMath Tier 4(v2)

第三者の結果(Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0)

出典のコストのある点が無いため、前線は出せません。

effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
GPT-6 AstraLow → Med+9.8pt—Epoch AI
GPT-6 AstraMed → High±0.0pt—Epoch AI
GPT-6 AstraHigh → Xhigh±0.0pt—Epoch AI
GPT-6 AstraXhigh → Max±0.0pt—Epoch AI

OTIS Mock AIME 2024-2025

第三者の結果(Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0)

出典のコストのある点が無いため、前線は出せません。

effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Haiku 5.5High → Xhigh+1.7pt—Epoch AI
Claude Fable 5Low → HighMed の値なし+2.2pt—Epoch AI
Claude Fable 5High → MaxXhigh の値なし-0.3pt—Epoch AI
Claude Opus 4.8Low → MaxMed・High・Xhigh の値なし+0.5pt—Epoch AI
Claude Opus 4.7Xhigh → Max-11.1pt—Epoch AI
Claude Sonnet 4.6Med → High-6.7pt—Epoch AI
Claude Sonnet 4.6High → Max-4.5pt—Epoch AI
GPT-5.6 SolNone → Low+26.7pt—Epoch AI
GPT-5.6 SolLow → MaxMed・High・Xhigh の値なし+4.4pt—Epoch AI
GPT-5.6 TerraNone → Low+35.6pt—Epoch AI
GPT-5.6 TerraLow → MaxMed・High・Xhigh の値なし+10.8pt—Epoch AI
GPT-5.6 LunaNone → Low+26.7pt—Epoch AI
GPT-5.6 LunaLow → MaxMed・High・Xhigh の値なし+31.7pt—Epoch AI
GPT-5.5None → Low+26.7pt—Epoch AI
Gemini 3.6 FlashMinimal → Low+2.2pt—Epoch AI
Gemini 3.6 FlashLow → HighMed の値なし+12.0pt—Epoch AI
Gemini 3.5 Flash-LiteMinimal → Low+8.9pt—Epoch AI
Gemini 3.5 Flash-LiteLow → HighMed の値なし+11.1pt—Epoch AI
Gemini 3.1 Flash-LiteMinimal → Low+6.7pt—Epoch AI
Gemini 3.1 Flash-LiteLow → HighMed の値なし+35.6pt—Epoch AI
Grok 4.6High → Xhigh+1.4pt—Epoch AI
Kimi K3Low → High+24.4pt—Epoch AI
Kimi K3High → Max+3.9pt—Epoch AI

SimpleQA Verified

第三者の結果(Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0)

出典のコストのある点が無いため、前線は出せません。

effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Sonnet 4.6High → Max-2.7pt—Epoch AI
Grok 4.6High → Xhigh-0.4pt—Epoch AI

Chess Puzzles(Epoch AI)

第三者の結果(Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0)

出典のコストのある点が無いため、前線は出せません。

effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Haiku 5.5Low → MaxMed・High・Xhigh の値なし+31.0pt—Epoch AI
Claude Fable 5Low → HighMed の値なし+14.0pt—Epoch AI
Claude Fable 5High → MaxXhigh の値なし+2.0pt—Epoch AI
Claude Opus 4.8Low → MaxMed・High・Xhigh の値なし+23.0pt—Epoch AI
Claude Opus 4.7Low → XhighMed・High の値なし+10.0pt—Epoch AI
Claude Opus 4.7Xhigh → Max+5.0pt—Epoch AI
Claude Sonnet 4.6Med → High+3.0pt—Epoch AI
Claude Sonnet 4.6High → Max+4.0pt—Epoch AI
GPT-5.6 SolNone → Low+24.0pt—Epoch AI
GPT-5.6 SolLow → MaxMed・High・Xhigh の値なし+28.0pt—Epoch AI
GPT-5.6 TerraNone → Low+22.0pt—Epoch AI
GPT-5.6 TerraLow → MaxMed・High・Xhigh の値なし+20.0pt—Epoch AI
GPT-5.6 LunaNone → Low+14.0pt—Epoch AI
GPT-5.6 LunaLow → MaxMed・High・Xhigh の値なし+26.0pt—Epoch AI
GPT-5.5None → Low+21.0pt—Epoch AI
GPT-5.5Low → XhighMed・High の値なし+33.0pt—Epoch AI
Gemini 3.6 FlashMinimal → Low+13.0pt—Epoch AI
Gemini 3.6 FlashLow → HighMed の値なし-2.0pt—Epoch AI
Gemini 3.5 Flash-LiteMinimal → Low+14.0pt—Epoch AI
Gemini 3.5 Flash-LiteLow → HighMed の値なし-5.0pt—Epoch AI
Grok 4.6High → Xhigh±0.0pt—Epoch AI
Kimi K3Low → High+19.0pt—Epoch AI
Kimi K3High → Max-1.0pt—Epoch AI

MirrorCode

第三者の結果(Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0)

出典のコストのある点が無いため、前線は出せません。

effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
Claude Opus 5.5High → MaxXhigh の値なし+16.7pt—Epoch AI
Claude Fable 5High → MaxXhigh の値なし+6.1pt—Epoch AI
GPT-6 AstraHigh → MaxXhigh の値なし+13.3pt—Epoch AI
GPT-5.5High → Xhigh+13.3pt—Epoch AI

τ²-bench airline (v1.0.1)

第三者の結果(τ²-bench v1.0.1(Sierra の実行))

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
GPT-5.2None52.5%$0.054Average cost in USD to run one trajectory in this domainSierra
GPT-5.2High83.0%$0.11Average cost in USD to run one trajectory in this domainSierra
Claude Opus 4.5High84.0%$0.40Average cost in USD to run one trajectory in this domainSierra

出典のコストの無い 4 点は前線の判定から外しています。

effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
GPT-5.2None → HighLow・Med の値なし+30.5pt×2.12Sierra

τ²-bench banking_knowledge (v1.0.1)

第三者の結果(τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 alltools))

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
GPT-5.4Xhigh39.4%$1.84Average cost in USD to run one trajectory in this domainSierra

出典のコストの無い 15 点は前線の判定から外しています。

第三者の結果(τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 terminal))

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
Claude Sonnet 4.5記載なし25.3%$4.05Average cost in USD to run one trajectory in this domainSierra

出典のコストの無い 2 点は前線の判定から外しています。

第三者の結果(τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 qwen_embeddings))

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
GPT-5.2None12.6%$0.24Average cost in USD to run one trajectory in this domainSierra

τ²-bench retail (v1.0.1)

第三者の結果(τ²-bench v1.0.1(Sierra の実行))

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
GPT-5.2None75.0%$0.048Average cost in USD to run one trajectory in this domainSierra
GPT-5.2High81.6%$0.072Average cost in USD to run one trajectory in this domainSierra

出典のコストの無い 4 点は前線の判定から外しています。

effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
GPT-5.2None → HighLow・Med の値なし+6.6pt×1.49Sierra

τ²-bench telecom (v1.0.1)

第三者の結果(τ²-bench v1.0.1(Sierra の実行))

前線上の点(安い順)
前線上のモデルeffortスコア出典のコスト(指標)計測・報告
GPT-5.2None57.2%$0.049Average cost in USD to run one trajectory in this domainSierra
GPT-5.2High89.7%$0.093Average cost in USD to run one trajectory in this domainSierra
Claude Opus 4.5High92.3%$0.72Average cost in USD to run one trajectory in this domainSierra

出典のコストの無い 4 点は前線の判定から外しています。

effort を上げたときの差分
モデルeffortスコアの差コストの倍率計測・報告
GPT-5.2None → HighLow・Med の値なし+32.4pt×1.88Sierra

指標別

ベンチマークをカテゴリの順に並べ、比較グループ(同じ出典か、同じ測定条件と確認できた出典の集まり)ごとに順位を付けます。「計測・報告」は出典に記載のある、値を計測または報告した主体で、公式発表でも他社のモデルの値を発表元が計測していることがあります。

エージェント型コーディングTerminal-Bench 4.0

公式の結果(Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Sonnet 5.5Max70.6%$12.54Cost per attempt前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
2Claude Opus 5.5Xhigh66.4%$7.35Cost per attempt前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
3Claude Opus 5.5Max64.8%$11.24Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
4Claude Opus 5.5High64.2%$3.88Cost per attempt前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
5Claude Sonnet 5.5Xhigh61.5%$5.30Cost per attempt—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
6GPT-6 AstraHigh57.9%$7.21Cost per attempt—OpenAIIntroducing Claude Opus 5.52026-11-12 取得
7Claude Opus 5.5Med57.6%$2.94Cost per attempt前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
7GPT-6 AstraXhigh57.6%$7.48Cost per attempt—OpenAIIntroducing Claude Opus 5.52026-11-12 取得
9GPT-6 AstraMax56.7%$10.35Cost per attempt—OpenAIIntroducing Claude Opus 5.52026-11-12 取得
10Claude Fable 5.1Max55.8%$19.50Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
11GPT-6 AstraMed53.9%$6.15Cost per attempt—OpenAIIntroducing Claude Opus 5.52026-11-12 取得
12Claude Opus 5Max52.3%$15.83Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
13Claude Fable 5.1Xhigh51.3%$15.80Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
14Claude Opus 5Xhigh50.6%$13.48Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
15GPT-6 AstraLow49.7%$4.95Cost per attempt—OpenAIIntroducing Claude Opus 5.52026-11-12 取得
16Claude Fable 5.1High49.4%$10.50Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
17Claude Opus 5High47.0%$10.64Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
18Claude Fable 5.1Med43.4%$7.80Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
19Claude Sonnet 5.5High43.0%$1.94Cost per attempt前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
20Claude Opus 5Med41.2%$7.00Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
21Claude Fable 5.1Low40.2%$5.70Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
22Claude Opus 5.5Low38.5%$1.29Cost per attempt前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
23GPT-5.6 SolMax37.3%$7.89Cost per attempt—OpenAIIntroducing Claude Sonnet 5.52026-11-12 取得
24Claude Sonnet 5.5Med28.8%$0.83Cost per attempt前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
25Claude Opus 5Low28.5%$4.25Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
25GPT-5.6 SolXhigh28.5%$5.39Cost per attempt—OpenAIIntroducing Claude Sonnet 5.52026-11-12 取得
27GPT-5.6 SolHigh26.1%$4.12Cost per attempt—OpenAIIntroducing Claude Sonnet 5.52026-11-12 取得
28GPT-5.6 SolMed20.9%$2.69Cost per attempt—OpenAIIntroducing Claude Sonnet 5.52026-11-12 取得
29Claude Sonnet 5.5Low20.0%$0.76Cost per attempt前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
30Claude Sonnet 5Max10.3%$11.62Cost per attempt—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
31GPT-5.6 SolLow7.9%$1.46Cost per attempt—OpenAIIntroducing Claude Sonnet 5.52026-11-12 取得
32Claude Sonnet 5Xhigh7.0%$9.95Cost per attempt—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
33Claude Sonnet 5High4.5%$8.20Cost per attempt—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
34Claude Sonnet 5Med4.4%$4.83Cost per attempt—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
35Claude Sonnet 5Low3.2%$3.78Cost per attempt—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得

公式の結果(Introducing Claude Haiku 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Sonnet 5.5Max70.6%$10.44Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
2Claude Sonnet 5.5Xhigh61.5%$4.34Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
3Claude Sonnet 5.5High43.0%$1.46Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
4Claude Haiku 5.5Max39.2%$2.64Cost per attempt—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
5Claude Haiku 5.5Xhigh31.5%$1.75Cost per attempt—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
6Claude Sonnet 5.5Med28.8%$0.68Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
7Claude Haiku 5.5High24.8%$1.04Cost per attempt—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
8Claude Haiku 5.5Med20.3%$0.68Cost per attempt—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
9Claude Sonnet 5.5Low20.0%$0.62Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
10GPT-6 Luna記載なし16.4%——記載なしIntroducing Claude Haiku 5.52026-11-12 取得
11Claude Haiku 5.5Low12.7%$0.42Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
12Claude Haiku 4.5記載なし0.0%$0.79Cost per attempt—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh65.2%——Artificial AnalysisGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh57.1%——Artificial AnalysisGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

公式の結果(Gemini 4 Argon model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 Argon記載なし57.4%——GoogleGemini 4 Argon model card2026-11-12 取得

公式の結果(Gemini 3.8 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.8 Flash記載なし19.1%——GoogleGemini 3.8 Flash model card2026-11-12 取得

公式の結果(Grok 4.7 (xAI news))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Grok 4.7Xhigh37.6%——xAIGrok 4.7 (xAI news)2026-11-12 取得
2Grok 4.6High20.3%——xAIGrok 4.7 (xAI news)2026-11-12 取得

公式の結果(Introducing Kimi K3.1 Pro)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax65.7%——Moonshot AIIntroducing Kimi K3.1 Pro2026-11-12 取得

公式の結果(Introducing Mistral Large 4)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Mistral Large 4記載なし28.3%——Mistral AIIntroducing Mistral Large 42026-11-12 取得

公式の結果(DeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1DeepSeek-V4.1-Flash記載なし31.2%——DeepSeekDeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai)2026-11-13 取得

公式の結果(DeepSeek-V4.1-Flash-1017 (Hugging Face model card, deepseek-ai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1DeepSeek-V4.1-Flash-1017記載なし42.3%——DeepSeekDeepSeek-V4.1-Flash-1017 (Hugging Face model card, deepseek-ai)2026-11-13 取得

エージェント型コーディングTerminal-Bench 3.0

公式の結果(GLM-5.3 (Z.ai docs))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GLM-5.3記載なし28.3%——Z.aiGLM-5.3 (Z.ai docs)2026-11-12 取得

公式の結果(Gemini 3.7 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.7 Flash記載なし14.9%——GoogleGemini 3.7 Flash model card2026-11-13 取得
2Gemini 3.6 Flash記載なし5.4%——GoogleGemini 3.7 Flash model card2026-11-13 取得

公式の結果(DeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1DeepSeek-V4.1-Flash記載なし30.0%——DeepSeekDeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai)2026-11-13 取得

エージェント型コーディングTerminal-Bench 2.1

公式の結果(Gemini 3.8 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.8 Flash記載なし89.4%——GoogleGemini 3.8 Flash model card2026-11-12 取得

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max88.3%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(GLM-5.3 (Hugging Face model card, zai-org))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GLM-5.3記載なし88.2%——Z.aiGLM-5.3 (Hugging Face model card, zai-org)2026-11-13 取得

公式の結果(Gemini 3.7 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.7 Flash記載なし85.8%——GoogleGemini 3.7 Flash model card2026-11-13 取得
2Gemini 3.6 Flash記載なし78.0%——GoogleGemini 3.7 Flash model card2026-11-13 取得

公式の結果(Gemini 3.5 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.5 Flash記載なし76.2%——GoogleGemini 3.5 Flash model card2026-11-13 取得
2Gemini 3 Flash Preview記載なし58.0%——GoogleGemini 3.5 Flash model card2026-11-13 取得

公式の結果(Qwen3.8-2.4T-A95B (Hugging Face model card, Qwen))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Qwen3.8-Max記載なし86.6%——Alibaba (Qwen)Qwen3.8-2.4T-A95B (Hugging Face model card, Qwen)2026-11-13 取得

公式の結果(DeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1DeepSeek-V4.1-Flash記載なし90.6%——DeepSeekDeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai)2026-11-13 取得

エージェント型コーディングTerminal-Bench-Science 0.1

公式の結果(Introducing Claude Opus 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GPT-6 Astra記載なし64.6%——OpenAIIntroducing Claude Opus 5.52026-11-12 取得
2Claude Opus 5.5Max58.7%——AnthropicIntroducing Claude Opus 5.52026-11-12 取得
3Claude Fable 5.1記載なし52.6%——AnthropicIntroducing Claude Opus 5.52026-11-12 取得
4Claude Opus 5記載なし29.0%——AnthropicIntroducing Claude Opus 5.52026-11-12 取得
5GPT-5.6 Sol記載なし22.4%——記載なしIntroducing Claude Opus 5.52026-11-12 取得

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Mythos 5.2記載なし75.3%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Claude Fable 5.2記載なし74.6%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3GPT-6 Astra記載なし68.1%——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
4Gemini 4 Carbon記載なし64.8%——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
5Claude Opus 5.5記載なし58.7%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
6Claude Fable 5.1記載なし52.6%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh64.8%——Artificial AnalysisGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh52.9%——Artificial AnalysisGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax38.1%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

エージェント型コーディングFrontierCode v1.1(main set)

公式の結果(Anthropic 発表の Accuracy vs Cost 図(Opus 5.5・Sonnet 5.5 ページ))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 5.5Med54.6%$0.80Cost per task前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
2Claude Opus 5.5Max54.4%$6.19Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
3Claude Opus 5.5High54.0%$1.09Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
4Claude Opus 5Med53.4%$4.61Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
5GPT-6 AstraMax53.3%$4.36Cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
6Claude Fable 5.1Low52.8%$2.47Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
7Claude Sonnet 5.5Xhigh52.1%$1.59Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
8Claude Opus 5.5Xhigh51.4%$2.25Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
9Claude Fable 5.1Med50.9%$3.28Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
9GPT-6 AstraHigh50.9%$2.85Cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
11GPT-6 AstraXhigh50.6%$3.10Cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
12Claude Fable 5.1High50.3%$5.27Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
12Claude Fable 5.1Max50.3%$12.82Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
14Claude Sonnet 5.5High49.4%$0.42Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
15GPT-6 SolMax49.3%$2.07Cost per task—記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
16GPT-6 AstraMed48.8%$2.28Cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
17Claude Fable 5.1Xhigh48.7%$9.27Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
18GPT-6 SolXhigh48.4%$1.32Cost per task—記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
19Claude Opus 5High48.0%$7.62Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
19Claude Opus 5Max48.0%$12.28Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
21GPT-6 SolHigh47.7%$1.04Cost per task—記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
22GPT-5.6 SolMax47.5%$4.85Cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
23Claude Opus 5.5Low47.3%$0.40Cost per task前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
24GPT-5.6 SolXhigh46.8%$3.88Cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
25Claude Sonnet 5.5Max46.2%$20.78Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
26GPT-6 SolMed45.9%$0.77Cost per task—記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
27GPT-6 AstraLow45.3%$1.59Cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
28GPT-5.6 SolHigh45.1%$3.25Cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
29Claude Opus 5Xhigh43.6%$8.99Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
30Claude Sonnet 5Xhigh42.7%$10.07Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
31Claude Sonnet 5Max42.4%$17.12Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
32Claude Opus 5Low42.0%$2.64Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
33GPT-5.6 SolMed39.9%$2.50Cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
34Claude Sonnet 5High39.4%$6.10Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
35GPT-6 SolLow37.3%$0.43Cost per task—記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
36Claude Sonnet 5.5Med36.5%$0.24Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
37GPT-5.6 SolLow35.4%$1.75Cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
38Claude Sonnet 5Med35.2%$3.81Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
39Claude Sonnet 5.5Low29.3%$0.19Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
40Claude Sonnet 5Low28.7%$2.39Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得

公式の結果(Introducing Claude Haiku 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Sonnet 5.5Xhigh52.1%——AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
2Claude Haiku 5.5記載なし46.4%——AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
3GPT-6 Luna記載なし42.4%——記載なしIntroducing Claude Haiku 5.52026-11-12 取得

エージェント型コーディングFrontierCode 2.0

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2Xhigh49.1%$14.19Cost per task前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Claude Fable 5.2Max45.6%$44.70Cost per task—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3Claude Opus 5.5Max45.5%$12.82Cost per task前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
4Claude Fable 5.2High45.3%$6.72Cost per task前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
5Claude Opus 5.5Xhigh43.4%$5.18Cost per task前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
6GPT-6 AstraMax42.8%$6.68Cost per task—記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
7GPT-6 AstraXhigh41.4%$4.54Cost per task前線記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
8Gemini 4 CarbonMed41.1%$5.16Cost per task—記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
9Gemini 4 CarbonHigh38.4%$5.83Cost per task—記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
10GPT-6 AstraHigh38.3%$4.28Cost per task前線記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
11Claude Fable 5.2Med37.6%$2.48Cost per task前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
12GPT-6 AstraMed36.8%$3.49Cost per task—記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
13Claude Opus 5.5High34.5%$2.42Cost per task前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
14Claude Fable 5.2Low34.2%$2.10Cost per task前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
15Claude Fable 5.1Xhigh32.4%$8.68Cost per task—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
15Gemini 4 CarbonLow32.4%$5.49Cost per task—記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
17Claude Fable 5.1Max31.7%$11.46Cost per task—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
18GPT-6 AstraLow31.1%$2.44Cost per task—記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
19Claude Opus 5.5Med30.1%$1.70Cost per task前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
20Claude Fable 5.1High25.7%$5.39Cost per task—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
21Claude Fable 5.1Med21.5%$3.89Cost per task—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
22Claude Fable 5.1Low20.6%$3.01Cost per task—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
23Claude Opus 5.5Low14.7%$0.75Cost per task前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

エージェント型コーディングFrontierSWE 2

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2記載なし69.9%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2GPT-6 Astra記載なし65.5%——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3Claude Opus 5.5記載なし62.3%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
4Gemini 4 Carbon記載なし61.2%——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
5Claude Fable 5.1記載なし56.3%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh61.2%——ProximalGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh55.0%——ProximalGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

エージェント型コーディングCursorBench 4.0

公式の結果(Introducing Claude Opus 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 5.5Max57.8%$13.43Cost per task前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
2Claude Opus 5.5High56.0%$3.97Cost per task前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
2Claude Opus 5.5Xhigh56.0%$6.99Cost per task前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
4Claude Opus 5.5Med52.5%$2.90Cost per task前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
5Claude Fable 5.1Max51.8%$17.28Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
6Claude Fable 5.1Xhigh51.6%$13.01Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
7Claude Fable 5.1High49.2%$9.08Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
8Claude Fable 5.1Med46.8%$7.05Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
9Claude Opus 5Max46.6%$11.95Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
10Claude Opus 5Xhigh46.1%$11.43Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
11Claude Fable 5.1Low45.1%$5.44Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
12Claude Opus 5High44.7%$9.00Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
13Claude Opus 5.5Low43.7%$1.18Cost per task前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
14Claude Opus 5Med43.3%$6.94Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
15GPT-5.6 SolMax41.7%$8.23Cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
16Claude Opus 5Low40.7%$4.87Cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
17GPT-5.6 SolXhigh37.7%$4.40Cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
18GPT-5.6 SolHigh35.7%$2.85Cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
19GPT-5.6 SolMed31.1%$1.77Cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
20GPT-5.6 SolLow24.6%$0.87Cost per task前線記載なしIntroducing Claude Opus 5.52026-11-12 取得

公式の結果(Introducing Claude Sonnet 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 5.5Max57.8%$13.43Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
2Claude Opus 5.5High56.0%$3.97Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
2Claude Opus 5.5Xhigh56.0%$6.98Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
4Claude Sonnet 5.5Max55.5%$9.67Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
5Claude Sonnet 5.5Xhigh53.1%$3.88Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
6Claude Opus 5.5Med52.5%$2.91Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
7Claude Sonnet 5.5High47.8%$1.67Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
8Claude Opus 5.5Low43.7%$1.17Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
9GPT-5.6 SolMax41.7%$8.23Cost per task—記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
10Claude Sonnet 5.5Med39.2%$0.70Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
11GPT-5.6 SolXhigh37.7%$4.40Cost per task—記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
12Claude Sonnet 5.5Low35.8%$0.50Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
13GPT-5.6 SolHigh35.7%$2.85Cost per task—記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
14Claude Sonnet 5Max34.1%$7.17Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
15Claude Sonnet 5Xhigh32.0%$4.55Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
16GPT-5.6 SolMed31.1%$1.77Cost per task—記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
17Claude Sonnet 5High30.8%$3.48Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
18Claude Sonnet 5Med28.0%$2.31Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
19GPT-5.6 SolLow24.6%$0.87Cost per task—記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
20Claude Sonnet 5Low24.1%$1.39Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得

公式の結果(Grok 4.7 (xAI news))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Grok 4.7Xhigh46.3%——xAIGrok 4.7 (xAI news)2026-11-12 取得
2Grok 4.6High40.4%——xAIGrok 4.7 (xAI news)2026-11-12 取得

エージェント型コーディングDeepSWE v1.1

公式の結果(Gemini 4 Argon model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 Argon記載なし77.9%——GoogleGemini 4 Argon model card2026-11-12 取得

公式の結果(Gemini 3.8 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.8 Flash記載なし73.7%——GoogleGemini 3.8 Flash model card2026-11-12 取得

公式の結果(Grok 4.7 (xAI news))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Grok 4.7High71.0%——xAIGrok 4.7 (xAI news)2026-11-12 取得
2Grok 4.6High65.2%——xAIGrok 4.7 (xAI news)2026-11-12 取得

公式の結果(GLM-5.3 (Z.ai docs))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GLM-5.3記載なし66.9%——Z.aiGLM-5.3 (Z.ai docs)2026-11-12 取得

公式の結果(Introducing Mistral Large 4)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Mistral Large 4記載なし61.7%——Mistral AIIntroducing Mistral Large 42026-11-12 取得

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax75.5%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max67.5%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(GLM-5.3-Flash (Z.ai docs))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GLM-5.3-Flash記載なし63.4%——Z.aiGLM-5.3-Flash (Z.ai docs)2026-11-13 取得

公式の結果(Gemini 3.7 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.7 FlashHigh65.3%——GoogleGemini 3.7 Flash model card2026-11-13 取得

公式の結果(Qwen3.8-2.4T-A95B (Hugging Face model card, Qwen))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Qwen3.8-Max記載なし56.6%——Alibaba (Qwen)Qwen3.8-2.4T-A95B (Hugging Face model card, Qwen)2026-11-13 取得

公式の結果(Qwen3.8-27B (Hugging Face model card, Qwen))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Qwen3.8-27B記載なし42.2%——Alibaba (Qwen)Qwen3.8-27B (Hugging Face model card, Qwen)2026-11-13 取得

公式の結果(DeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1DeepSeek-V4.1-Flash記載なし74.2%——DeepSeekDeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai)2026-11-13 取得

公式の結果(DeepSeek-V4.1-Flash-1017 (Hugging Face model card, deepseek-ai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1DeepSeek-V4.1-Flash-1017記載なし77.0%——DeepSeekDeepSeek-V4.1-Flash-1017 (Hugging Face model card, deepseek-ai)2026-11-13 取得

エージェント型コーディングSWE-sweep

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh9.6%——GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

エージェント型コーディングSWE-Bench Pro

公式の結果(Gemini 3.5 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.5 Flash記載なし55.1%——GoogleGemini 3.5 Flash model card2026-11-13 取得
2Gemini 3.1 Pro Preview記載なし54.2%——GoogleGemini 3.5 Flash model card2026-11-13 取得
3Gemini 3 Flash Preview記載なし49.6%——GoogleGemini 3.5 Flash model card2026-11-13 取得

公式の結果(Qwen3.8-2.4T-A95B (Hugging Face model card, Qwen))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Qwen3.8-Max記載なし67.7%——Alibaba (Qwen)Qwen3.8-2.4T-A95B (Hugging Face model card, Qwen)2026-11-13 取得

公式の結果(Qwen3.8-27B (Hugging Face model card, Qwen))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Qwen3.8-27B記載なし61.7%——Alibaba (Qwen)Qwen3.8-27B (Hugging Face model card, Qwen)2026-11-13 取得

エージェント型コーディングSWE-Marathon v1.1

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax51.7%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(GLM-5.3 (Hugging Face model card, zai-org))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GLM-5.3Max42.5%——Z.aiGLM-5.3 (Hugging Face model card, zai-org)2026-11-13 取得

エージェント型コーディングSWE-Marathon(版の記載なし)

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max42.0%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

エージェント型コーディングProgramBench(Almost Solved)

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax31.5%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(DeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1DeepSeek-V4.1-Flash記載なし20.3%——DeepSeekDeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai)2026-11-13 取得

エージェント型コーディングNL2Repo-Bench

公式の結果(GLM-5.3 (Hugging Face model card, zai-org))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GLM-5.3記載なし58——Z.aiGLM-5.3 (Hugging Face model card, zai-org)2026-11-13 取得

公式の結果(Qwen3.8-2.4T-A95B (Hugging Face model card, Qwen))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Qwen3.8-Max記載なし55.9——Alibaba (Qwen)Qwen3.8-2.4T-A95B (Hugging Face model card, Qwen)2026-11-13 取得

公式の結果(Qwen3.8-27B (Hugging Face model card, Qwen))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Qwen3.8-27B記載なし42.3——Alibaba (Qwen)Qwen3.8-27B (Hugging Face model card, Qwen)2026-11-13 取得

公式の結果(DeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1DeepSeek-V4.1-Flash記載なし64——DeepSeekDeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai)2026-11-13 取得

エージェント型コーディングMLS-Bench-Lite

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max48.3%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Qwen3.8-2.4T-A95B (Hugging Face model card, Qwen))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Qwen3.8-Max記載なし41.0%——Alibaba (Qwen)Qwen3.8-2.4T-A95B (Hugging Face model card, Qwen)2026-11-13 取得

コーディングCWE-bench v1

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh71.0%——Collinear AI / Artificial AnalysisGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh68.0%——Collinear AI / Artificial AnalysisGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

コーディングVibe Code Bench

公式の結果(Gemini 4 Argon model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 Argon記載なし91.9%——GoogleGemini 4 Argon model card2026-11-12 取得

コーディングSWE-Atlas-QnA

公式の結果(Introducing Mistral Large 4)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Mistral Large 4記載なし59.4%——Mistral AIIntroducing Mistral Large 42026-11-12 取得

コーディングMirrorCode

第三者の結果(Epoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.0)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Sonnet 5.5Max97.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
2Claude Opus 5.5Max95.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
3Claude Haiku 5.5Max78.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
3Claude Opus 5.5High78.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
5GPT-6.1 TerraMax76.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
6Claude Fable 5Max70.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
7Claude Opus 4.8Max65.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
8DeepSeek-V4.1-Flash記載なし64.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
9Claude Fable 5High63.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
10GPT-6.1 SolMax61.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
11GPT-6 AstraMax60.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
12Gemini 4 ArgonHigh58.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
13Grok 4.7Xhigh54.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
13Kimi K3Max54.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
15Qwen3.8-MaxXhigh52.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
16GPT-6 SolMax48.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
17GPT-6 AstraHigh46.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
18Grok 4.6Xhigh41.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
18GPT-5.6 TerraMax41.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
18Gemini 3.7 FlashHigh41.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
21Gemini 3.8 FlashHigh38.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
22Claude Opus 4.6Max33.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
22DeepSeek-V4-Pro-0813Max33.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
24Claude Opus 4.7High31.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
25Mistral Large 4High27.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
26GPT-6 LunaMax26.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
27GPT-5.5Xhigh23.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
27Gemini 3.6 FlashHigh23.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
29GPT-5.6 LunaMax21.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
30GPT-5.6 SolHigh20.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
31GPT-5.5High10.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得
32Gemini 3.1 Pro PreviewHigh8.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, mirrorcode.csv) CC BY 4.02026-11-12 取得

コーディングLMArena Text Coding(style control なし)

第三者の結果(lmarena-ai/leaderboard-dataset text/latest-00000-of-00001.parquet (CC BY 4.0))

推論Humanity's Last Exam(ツールあり)

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2Max74.2%$4.51Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Claude Fable 5.2Xhigh73.4%$1.48Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3Claude Fable 5.2High71.5%$0.63Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
4Claude Fable 5.2Med68.6%$0.21Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
5Claude Fable 5.2Low68.4%$0.19Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
6Claude Opus 5.5Max67.7%$2.18Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
7Claude Opus 5.5Xhigh66.4%$0.69Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
8Claude Fable 5.1Max65.6%$3.77Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
9Claude Fable 5.1Xhigh65.2%$2.70Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
10Claude Opus 5.5High63.9%$0.29Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
11Claude Fable 5.1High63.1%$1.37Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
12Claude Opus 5.5Med63.0%$0.17Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
13Claude Fable 5.1Med62.4%$0.85Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
14Claude Fable 5.1Low60.2%$0.59Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
15Claude Opus 5.5Low57.4%$0.08Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
16GPT-6 Astra記載なし57.2%——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

公式の結果(Introducing Claude Opus 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 5.5Max67.7%——AnthropicIntroducing Claude Opus 5.52026-11-12 取得
2Claude Fable 5.1記載なし65.6%——AnthropicIntroducing Claude Opus 5.52026-11-12 取得
3Claude Opus 5記載なし63.6%——AnthropicIntroducing Claude Opus 5.52026-11-12 取得
4GPT-6 Astra記載なし57.2%——記載なしIntroducing Claude Opus 5.52026-11-12 取得

公式の結果(Introducing Claude Sonnet 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 5.5記載なし67.7%——AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
2Claude Sonnet 5.5記載なし64.5%——AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
3Claude Sonnet 5記載なし54.9%——AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得

公式の結果(Introducing Claude Haiku 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Sonnet 5.5記載なし64.5%——AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
2Claude Haiku 5.5記載なし57.4%——AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
3Claude Haiku 4.5記載なし18.7%——AnthropicIntroducing Claude Haiku 5.52026-11-12 取得

公式の結果(GLM-5.3 (Hugging Face model card, zai-org))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GLM-5.3記載なし62.5%——Z.aiGLM-5.3 (Hugging Face model card, zai-org)2026-11-13 取得

公式の結果(Introducing Claude Opus 4.6)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 4.6Max53.0%——AnthropicIntroducing Claude Opus 4.62026-11-13 取得

推論Humanity's Last Exam(ツールなし)

公式の結果(Introducing Claude Haiku 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Sonnet 5.5Max56.9%$0.89Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
2Claude Sonnet 5.5Xhigh53.0%$0.20Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
3Claude Sonnet 5.5High47.9%$0.08Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
4Claude Haiku 5.5Max45.9%$0.21Cost per attempt—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
5Claude Haiku 5.5Xhigh44.1%$0.06Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
6Claude Sonnet 5.5Med43.2%$0.03Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
7Claude Sonnet 5.5Low41.4%$0.03Cost per attempt—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
8Claude Haiku 5.5High39.7%$0.02Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
9Claude Haiku 5.5Med35.5%$0.007Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
10Claude Haiku 5.5Low30.9%$0.003Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
11Claude Haiku 4.5記載なし10.2%$0.09Cost per attempt—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得

推論HLE-Verified

公式の結果(Gemini 3.8 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.8 Flash記載なし54.9%——GoogleGemini 3.8 Flash model card2026-11-12 取得

公式の結果(Gemini 3.7 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.7 Flash記載なし53.6%——GoogleGemini 3.7 Flash model card2026-11-13 取得
2Gemini 3.6 Flash記載なし51.2%——GoogleGemini 3.7 Flash model card2026-11-13 取得

推論GPQA Diamond

第三者の結果(Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.0)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh96.5%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
2Gemini 4 ArgonHigh96.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
3GPT-6 AstraMax95.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
4Claude Sonnet 5.5Max95.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
5GPT-6.1 SolMax95.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
5Gemini 3.8 FlashHigh95.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
7Gemini 3.7 FlashHigh94.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
8Gemini 3.1 Pro PreviewHigh94.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
9GPT-6.1 TerraMax94.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
9GPT-6 SolMax94.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
11Kimi K3.1 ProMax94.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
12Gemini 3.6 FlashHigh94.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
13Gemini 3.1 Pro Preview記載なし94.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
14Grok 4.6High94.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
15Claude Fable 5.2Max93.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
16GPT-5.6 SolMax93.5%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
17GPT-5.6 TerraMax93.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
18Grok 4.6Xhigh93.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
18DeepSeek-V4.1-Flash-1017Max93.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
20Kimi K3Max93.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
21Grok 4.7Xhigh92.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
22Qwen3.8-MaxXhigh92.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
23Kimi K3High91.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
24DeepSeek-V4-Pro-0813Max91.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
25GPT-5.6 LunaMax91.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
26Claude Opus 4.8Max91.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
27GLM-5.3Max90.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
28GPT-5.5Low90.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
29Claude Opus 5.5Max90.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
30GPT-6 LunaMax90.5%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
31Claude Opus 4.7Xhigh90.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
31GLM-5.3-FlashMax90.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
33GPT-5.6 SolLow89.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
34DeepSeek-V4.1-Flash記載なし89.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
35Claude Haiku 5.5Max89.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
36Gemini 3 Flash PreviewHigh89.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
37Claude Opus 4.6Max88.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
37Claude Opus 4.8Low88.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
39Mistral Large 4High88.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
40GPT-5.6 TerraLow87.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
41Gemini 3.6 FlashLow86.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
41Claude Opus 4.7Max86.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
43Gemini 3.6 FlashMinimal85.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
43Claude Fable 5Max85.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
45Claude Opus 4.8記載なし85.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
46Gemini 2.5 Pro記載なし85.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
47Kimi K3Low84.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
48Gemini 3.5 Flash-LiteHigh83.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
48Claude Fable 5High83.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
48Claude Sonnet 4.6High83.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
48Claude Sonnet 4.6Med83.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
52Gemini 3 Flash Preview記載なし83.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
53GPT-5.6 SolNone82.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
54GPT-5.6 LunaLow82.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
55Gemini 3.1 Flash-LiteHigh81.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
56Claude Opus 4.5記載なし80.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
57Claude Sonnet 4.6Max78.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
57Claude Fable 5Low78.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
59GPT-5.5None77.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
59GPT-5.6 TerraNone77.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
61Gemini 3.5 Flash-LiteLow75.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
62Gemini 3.5 Flash-LiteMinimal74.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
62Gemini 3.1 Flash-LiteLow74.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
64Gemini 3.1 Flash-LiteMinimal73.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
64Claude Sonnet 4.5記載なし73.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得
66GPT-5.6 LunaNone63.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv) CC BY 4.02026-11-12 取得

推論Chess Puzzles(Epoch AI)

第三者の結果(Epoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.0)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2Max79.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
2Gemini 4 CarbonHigh78.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
3GPT-6 AstraMax74.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
4GPT-6.1 SolMax68.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
5Claude Opus 5.5Max67.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
6Gemini 3.7 FlashHigh66.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
7GPT-5.5Xhigh63.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
8Gemini 4 ArgonHigh59.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
8Claude Sonnet 5.5Max59.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
8GPT-5.6 SolMax59.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
11Gemini 3.8 FlashHigh58.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
12GPT-6 SolMax56.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
13Claude Haiku 5.5Max55.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
14Claude Opus 4.8Max52.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
15Gemini 3.1 Pro PreviewHigh51.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
16DeepSeek-V4.1-Flash記載なし48.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
17GPT-5.6 TerraMax47.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
17Claude Fable 5Max47.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
19Qwen3.8-MaxXhigh46.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
19DeepSeek-V4.1-Flash-1017Max46.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
19DeepSeek-V4-Pro-0813Max46.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
22Kimi K3.1 ProMax45.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
22Claude Fable 5High45.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
24GPT-5.6 LunaMax43.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
25Gemini 3.6 FlashLow42.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
26Claude Opus 4.7Max41.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
26Grok 4.7Xhigh41.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
28Gemini 3 Flash PreviewHigh40.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
28Gemini 3.6 FlashHigh40.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
30GLM-5.3Max37.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
30Kimi K3High37.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
32GPT-6 LunaMax36.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
32Claude Opus 4.7Xhigh36.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
32Kimi K3Max36.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
35Mistral Large 4High33.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
35Grok 4.6Xhigh33.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
35Grok 4.6High33.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
38GPT-5.6 SolLow31.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
38Claude Fable 5Low31.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
40GPT-5.5Low30.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
41Claude Opus 4.8Low29.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
41Gemini 3.6 FlashMinimal29.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
43GPT-5.6 TerraLow27.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
44Claude Opus 4.7Low26.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
45Claude Haiku 5.5Low24.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
46Claude Sonnet 4.6Max23.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
47Claude Opus 4.6Max21.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
47Gemini 3.5 Flash-LiteLow21.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
47GLM-5.3-FlashMax21.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
50Claude Sonnet 4.6High19.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
50Claude Opus 4.8記載なし19.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
52Kimi K3Low18.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
53GPT-5.6 LunaLow17.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
54Claude Sonnet 4.6Med16.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
54Gemini 3.5 Flash-LiteHigh16.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
56GPT-5.5None9.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
57GPT-5.6 SolNone7.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
57Gemini 3.5 Flash-LiteMinimal7.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
59Claude Sonnet 4.5記載なし6.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
60GPT-5.6 TerraNone5.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
60Claude Opus 4.5記載なし5.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得
62GPT-5.6 LunaNone3.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, chess_puzzles.csv) CC BY 4.02026-11-12 取得

数学RiemannBench

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 ArgonHigh76.0%——Surge AIGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 CarbonHigh72.0%——Surge AIGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

数学FrontierMath Tiers 1-3(v2)

第三者の結果(Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.0)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2Max95.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
2GPT-6.1 SolMax93.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
2GPT-6 AstraMax93.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
4Claude Opus 5.5Max91.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
5GPT-6.1 TerraMax90.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
6Gemini 4 CarbonHigh90.5%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
7GPT-6 SolMax89.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
8GPT-5.6 SolMax89.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
9Claude Sonnet 5.5Max88.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
10Claude Fable 5Max87.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
11Gemini 4 ArgonHigh86.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
11GPT-5.6 TerraMax86.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
13GPT-5.5Xhigh85.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
14GPT-5.6 LunaMax82.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
15Claude Opus 4.8Max80.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
16GPT-6 LunaMax79.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
17Kimi K3.1 ProMax78.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
18Claude Haiku 5.5Max75.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
19Qwen3.8-MaxXhigh74.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
20Kimi K3Max72.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
21Gemini 3.7 FlashHigh71.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
22DeepSeek-V4.1-Flash-1017Max71.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
23Claude Opus 4.7Max70.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
24GLM-5.3Max68.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
25Gemini 3.8 FlashHigh68.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
26DeepSeek-V4.1-Flash記載なし67.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
27Grok 4.6Xhigh66.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
27Claude Opus 4.6Max66.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
29DeepSeek-V4-Pro-0813Max64.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
30Gemini 3.1 Pro Preview記載なし59.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
31Gemini 3.6 FlashHigh59.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
32Mistral Large 4High58.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
33GLM-5.3-FlashMax55.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
34Grok 4.7Xhigh53.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
35Gemini 3 Flash Preview記載なし51.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
36GPT-5.6 LunaLow41.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
37GPT-5.6 LunaNone39.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
38Gemini 3.1 Flash-LiteHigh27.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
39Gemini 3.5 Flash-LiteHigh26.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
40Gemini 2.5 Pro記載なし24.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
41Gemini 3.1 Flash-LiteLow22.5%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得
42Gemini 3.1 Flash-LiteMinimal21.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tiers_1_3_v2.csv) CC BY 4.02026-11-12 取得

数学FrontierMath Tier 4(v2)

第三者の結果(Epoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.0)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2Max100.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
1GPT-6.1 SolMax100.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
3GPT-6 AstraHigh97.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
3GPT-6 AstraXhigh97.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
3GPT-6 AstraMax97.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
6GPT-6 AstraMed97.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
7Gemini 4 CarbonHigh95.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
8Claude Opus 5.5Max95.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
9GPT-6.1 TerraMax92.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
10Claude Fable 5Max90.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
11GPT-6 SolMax90.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
12GPT-6 AstraLow87.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
13GPT-6 Astra記載なし82.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
13GPT-5.6 SolMax82.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
15Claude Sonnet 5.5Max80.5%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
16Gemini 4 ArgonHigh78.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
17GPT-5.5Xhigh72.5%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
18GPT-5.6 TerraMax70.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
19Kimi K3.1 ProMax65.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
20GPT-5.6 LunaMax61.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
21GPT-6 LunaMax56.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
21Claude Opus 4.8Max56.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
23Claude Haiku 5.5Max46.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
23Qwen3.8-MaxXhigh46.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
23DeepSeek-V4.1-Flash-1017Max46.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
26Kimi K3Max39.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
27Gemini 3.7 FlashHigh36.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
28Grok 4.6Xhigh31.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
28Claude Opus 4.7Max31.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
30GLM-5.3Max29.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
31Mistral Large 4High26.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
31Gemini 3.1 Pro Preview記載なし26.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
31Claude Opus 4.6Max26.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
31DeepSeek-V4-Pro-0813Max26.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
31DeepSeek-V4.1-Flash記載なし26.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
36Gemini 3.8 FlashHigh22.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
36Gemini 3.6 FlashHigh22.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
38Grok 4.7Xhigh17.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
38Gemini 3 Flash Preview記載なし17.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
38GLM-5.3-FlashMax17.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
41Gemini 3.5 Flash-LiteHigh0.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得
41Gemini 2.5 Pro記載なし0.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, frontiermath_tier_4_v2.csv) CC BY 4.02026-11-12 取得

数学OTIS Mock AIME 2024-2025

第三者の結果(Epoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.0)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh100.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
1GPT-6.1 TerraMax100.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
1Gemini 4 ArgonHigh100.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
1GPT-6.1 SolMax100.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
1Claude Sonnet 5.5Max100.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
1Claude Opus 5.5Max100.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
1GPT-6 AstraMax100.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
1Kimi K3.1 ProMax100.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
1GPT-6 SolMax100.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
1Claude Fable 5High100.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
1GPT-5.6 SolMax100.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
12Claude Fable 5.2Max99.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
12GPT-5.6 TerraMax99.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
12Claude Fable 5Max99.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
15Qwen3.8-MaxXhigh99.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
15DeepSeek-V4.1-Flash-1017Max99.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
17Grok 4.6Xhigh99.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
18Claude Haiku 5.5Xhigh98.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
18GPT-6 LunaMax98.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
18Gemini 3.8 FlashHigh98.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
21DeepSeek-V4-Pro-0813Max98.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
22Mistral Large 4High98.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
22GPT-5.6 LunaMax98.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
22Claude Opus 4.8Max98.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
22DeepSeek-V4.1-Flash記載なし98.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
26Grok 4.7Xhigh98.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
27Claude Opus 4.7Xhigh97.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
28Grok 4.6High97.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
28Claude Opus 4.8Low97.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
28Claude Fable 5Low97.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
31Claude Haiku 5.5High97.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
31Gemini 3.7 FlashHigh97.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
31Kimi K3Max97.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
34Gemini 3.1 Pro Preview記載なし95.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
35GPT-5.6 SolLow95.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
35Gemini 3.1 Pro PreviewHigh95.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
35Gemini 3 Flash PreviewHigh95.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
38Gemini 3.6 FlashHigh94.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
39GLM-5.3-FlashMax93.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
40Kimi K3High93.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
41Gemini 3 Flash Preview記載なし92.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
42GLM-5.3Max91.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
42Claude Opus 4.6Max91.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
44GPT-5.6 TerraLow88.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
45Claude Opus 4.7Max86.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
46Gemini 2.5 Pro記載なし84.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
47GPT-5.5Low84.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
47Claude Opus 4.8記載なし84.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
49Gemini 3.6 FlashLow82.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
49Claude Sonnet 4.6Med82.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
51Gemini 3.6 FlashMinimal80.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
51Gemini 3.1 Flash-LiteHigh80.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
53Claude Sonnet 4.6High75.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
54Gemini 3.5 Flash-LiteHigh71.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
54Claude Sonnet 4.6Max71.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
56GPT-5.6 SolNone68.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
56Kimi K3Low68.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
58GPT-5.6 LunaLow66.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
59Gemini 3.5 Flash-LiteLow60.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
60GPT-5.5None57.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
61GPT-5.6 TerraNone53.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
62Gemini 3.5 Flash-LiteMinimal51.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
63Claude Opus 4.5記載なし48.1%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
64Gemini 3.1 Flash-LiteLow44.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
65GPT-5.6 LunaNone40.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
66Gemini 3.1 Flash-LiteMinimal37.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得
67Claude Sonnet 4.5記載なし35.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, otis_mock_aime_2024_2025.csv) CC BY 4.02026-11-12 取得

数学MathVision

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max94.3%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Qwen3.8-27B (Hugging Face model card, Qwen))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Qwen3.8-27B記載なし90.0%——Alibaba (Qwen)Qwen3.8-27B (Hugging Face model card, Qwen)2026-11-13 取得

数学MathVision(ツールあり)

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max97.8%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Qwen3.8-27B (Hugging Face model card, Qwen))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Qwen3.8-27B記載なし94.6%——Alibaba (Qwen)Qwen3.8-27B (Hugging Face model card, Qwen)2026-11-13 取得

知識GDPval-AA v2.1

公式の結果(Introducing Claude Opus 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 5.5Max1,846$8.92Estimated cost per task前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
2Claude Opus 5.5Xhigh1,820$4.21Estimated cost per task前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
3Claude Fable 5.1Max1,735$9.59Estimated cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
4Claude Fable 5.1Xhigh1,721$7.09Estimated cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
5Claude Opus 5Max1,708$6.76Estimated cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
6Claude Opus 5.5High1,692$1.54Estimated cost per task前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
7Claude Opus 5Xhigh1,676$4.96Estimated cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
8Claude Fable 5.1High1,617$3.43Estimated cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
9GPT-5.6 SolMax1,588$2.81Estimated cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
10Claude Opus 5High1,581$3.03Estimated cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
11Claude Opus 5.5Med1,576$0.86Estimated cost per task前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
12GPT-5.6 SolXhigh1,548$1.70Estimated cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
13GPT-6 AstraMax1,542$4.53Estimated cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
14Claude Fable 5.1Med1,536$2.17Estimated cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
15GPT-6 AstraXhigh1,516$3.04Estimated cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
16GPT-6 AstraHigh1,485$2.43Estimated cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
17GPT-5.6 SolHigh1,480$1.11Estimated cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
18Claude Opus 5Med1,476$1.36Estimated cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
19GPT-6 AstraMed1,468$1.82Estimated cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
20Claude Fable 5.1Low1,450$1.41Estimated cost per task—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
21GPT-5.6 SolMed1,403$0.60Estimated cost per task前線記載なしIntroducing Claude Opus 5.52026-11-12 取得
22GPT-6 AstraLow1,366$0.85Estimated cost per task—記載なしIntroducing Claude Opus 5.52026-11-12 取得
23Claude Opus 5Low1,294$0.58Estimated cost per task前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
24GPT-5.6 SolLow1,289$0.27Estimated cost per task前線記載なしIntroducing Claude Opus 5.52026-11-12 取得
25Claude Opus 5.5Low1,224$0.21Estimated cost per task前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得

公式の結果(Introducing Claude Haiku 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Sonnet 5.5Max1,840$6.78Cost per task前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
2Claude Sonnet 5.5Xhigh1,731$1.88Cost per task前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
3Claude Haiku 5.5Max1,620$0.87Cost per task前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
4Claude Sonnet 5.5High1,551$0.62Cost per task前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
5Claude Haiku 5.5Xhigh1,513$0.27Cost per task前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
6GPT-6 LunaMax1,437$0.09Cost per task前線記載なしIntroducing Claude Haiku 5.52026-11-12 取得
7Claude Haiku 5.5High1,420$0.09Cost per task—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
8GPT-6 LunaXhigh1,364$0.05Cost per task前線記載なしIntroducing Claude Haiku 5.52026-11-12 取得
9GPT-6 LunaHigh1,344$0.03Cost per task前線記載なしIntroducing Claude Haiku 5.52026-11-12 取得
10Claude Sonnet 5.5Med1,324$0.27Cost per task—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
11Claude Haiku 5.5Med1,277$0.03Cost per task—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
12GPT-6 LunaMed1,262$0.02Cost per task前線記載なしIntroducing Claude Haiku 5.52026-11-12 取得
13Claude Sonnet 5.5Low1,179$0.22Cost per task—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
14Claude Haiku 5.5Low1,125$0.01Cost per task前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
15GPT-6 LunaLow1,036$0.004Cost per task前線記載なしIntroducing Claude Haiku 5.52026-11-12 取得
16Claude Haiku 4.5記載なし735$0.24Cost per task—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得

公式の結果(Introducing Claude Sonnet 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 5.5記載なし1,846——Artificial AnalysisIntroducing Claude Sonnet 5.52026-11-12 取得
2Claude Sonnet 5.5記載なし1,844——Artificial AnalysisIntroducing Claude Sonnet 5.52026-11-12 取得
3GPT-6 Sol記載なし1,487——Artificial AnalysisIntroducing Claude Sonnet 5.52026-11-12 取得
4Claude Sonnet 5記載なし1,449——Artificial AnalysisIntroducing Claude Sonnet 5.52026-11-12 取得

知識GDPval-AA v2

公式の結果(Gemini 3.8 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.8 Flash記載なし1,545——GoogleGemini 3.8 Flash model card2026-11-12 取得

知識AA-Briefcase v1.1(Elo)

公式の結果(Introducing Claude Sonnet 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 5.5Max1,822$21.05Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
2Claude Sonnet 5.5Max1,811$29.19Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
3Claude Opus 5.5Xhigh1,780$12.27Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
4Claude Sonnet 5.5Xhigh1,746$9.63Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
5Claude Opus 5.5High1,705$6.27Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
6Claude Opus 5.5Med1,642$4.40Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
7Claude Sonnet 5.5High1,634$3.95Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
8GPT-6 SolMax1,483$2.67Cost per task前線記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
9Claude Sonnet 5.5Med1,461$1.64Cost per task前線AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
10GPT-6 SolXhigh1,364$1.19Cost per task前線記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
11Claude Sonnet 5Max1,359$14.43Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
12GPT-6 SolHigh1,289$0.63Cost per task前線記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
13Claude Opus 5.5Low1,285$1.15Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
14Claude Sonnet 5Xhigh1,274$7.56Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
15Claude Sonnet 5.5Low1,264$0.87Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
16Claude Sonnet 5High1,177$3.76Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
17GPT-6 SolMed1,142$0.34Cost per task前線記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
18Claude Sonnet 5Med1,056$1.73Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
19Claude Sonnet 5Low923$0.82Cost per task—AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
20GPT-6 SolLow905$0.12Cost per task前線記載なしIntroducing Claude Sonnet 5.52026-11-12 取得

公式の結果(Introducing Claude Haiku 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Sonnet 5.5記載なし1,824——AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
2Claude Haiku 5.5記載なし1,578——AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
3GPT-6 Luna記載なし1,336——記載なしIntroducing Claude Haiku 5.52026-11-12 取得
4Claude Haiku 4.5記載なし614——AnthropicIntroducing Claude Haiku 5.52026-11-12 取得

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 5.5記載なし1,800——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Claude Fable 5.2記載なし1,785——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3Claude Fable 5.1記載なし1,669——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
4Gemini 4 Carbon記載なし1,611——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
5GPT-6 Astra記載なし1,569——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

公式の結果(Grok 4.7 (xAI news))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Grok 4.7Xhigh1,657——xAIGrok 4.7 (xAI news)2026-11-12 取得
2Grok 4.6High1,546——xAIGrok 4.7 (xAI news)2026-11-12 取得

知識AA-Briefcase(Elo・版の記載なし)

公式の結果(Introducing Mistral Large 4)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Mistral Large 4記載なし1,393——Mistral AIIntroducing Mistral Large 42026-11-12 取得

知識AA-Briefcase v1.1(rubric)

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2記載なし66.1%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Gemini 4 Carbon記載なし61.8%——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3Claude Fable 5.1記載なし61.5%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
4Claude Opus 5.5記載なし61.2%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
5GPT-6 Astra記載なし52.0%——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

知識Remote Labor Index(RLI)

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2High41.3%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Claude Opus 5.5記載なし21.3%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3GPT-6 Astra記載なし20.8%——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
4Claude Fable 5.1記載なし17.9%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

知識HealthBench Professional(長さ調整スコア)

公式の結果(GPT-6 Astra System Card (Table 6. HealthBench Evaluations))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GPT-6 Astra記載なし64.7——OpenAIGPT-6 Astra System Card (Table 6. HealthBench Evaluations)2026-11-12 取得
2GPT-5.6 Sol記載なし60.5——OpenAIGPT-6 Astra System Card (Table 6. HealthBench Evaluations)2026-11-12 取得
3GPT-5.6 Terra記載なし57.7——OpenAIGPT-6 Astra System Card (Table 6. HealthBench Evaluations)2026-11-12 取得
4GPT-5.6 Luna記載なし55.7——OpenAIGPT-6 Astra System Card (Table 6. HealthBench Evaluations)2026-11-12 取得

知識HealthBench Hard(長さ調整スコア)

公式の結果(GPT-6 Astra System Card (Table 6. HealthBench Evaluations))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GPT-6 Astra記載なし36.6——OpenAIGPT-6 Astra System Card (Table 6. HealthBench Evaluations)2026-11-12 取得
2GPT-5.6 Sol記載なし33.1——OpenAIGPT-6 Astra System Card (Table 6. HealthBench Evaluations)2026-11-12 取得
3GPT-5.6 Terra記載なし32.7——OpenAIGPT-6 Astra System Card (Table 6. HealthBench Evaluations)2026-11-12 取得
4GPT-5.6 Luna記載なし32——OpenAIGPT-6 Astra System Card (Table 6. HealthBench Evaluations)2026-11-12 取得

知識HealthBench Professional

公式の結果(Grok 4.7 (xAI news))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Grok 4.7Xhigh56.7%——xAIGrok 4.7 (xAI news)2026-11-12 取得
2Grok 4.6High48.5%——xAIGrok 4.7 (xAI news)2026-11-12 取得

知識LABBench 2

公式の結果(Gemini 4 Argon model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 Argon記載なし88.8%——GoogleGemini 4 Argon model card2026-11-12 取得

公式の結果(Gemini 3.8 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.8 Flash記載なし86.2%——GoogleGemini 3.8 Flash model card2026-11-12 取得

公式の結果(Gemini 3.7 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.7 Flash記載なし82.1%——GoogleGemini 3.7 Flash model card2026-11-13 取得
2Gemini 3.6 Flash記載なし76.1%——GoogleGemini 3.7 Flash model card2026-11-13 取得

知識GDP.PDF

公式の結果(Gemini 3.8 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.8 Flash記載なし35.0%——GoogleGemini 3.8 Flash model card2026-11-12 取得

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax34.6%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Gemini 3.7 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.7 Flash記載なし34.0%——GoogleGemini 3.7 Flash model card2026-11-13 取得
2Gemini 3.6 Flash記載なし22.0%——GoogleGemini 3.7 Flash model card2026-11-13 取得

知識BioMysteryBench(Human Solvable)

公式の結果(Gemini 3.8 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.8 Flash記載なし88.8%——GoogleGemini 3.8 Flash model card2026-11-12 取得

公式の結果(Gemini 3.7 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.7 Flash記載なし87.1%——GoogleGemini 3.7 Flash model card2026-11-13 取得
2Gemini 3.6 Flash記載なし80.6%——GoogleGemini 3.7 Flash model card2026-11-13 取得

知識BioMysteryBench(Human Difficult)

公式の結果(Gemini 3.8 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.8 Flash記載なし56.5%——GoogleGemini 3.8 Flash model card2026-11-12 取得

公式の結果(Gemini 3.7 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.7 Flash記載なし43.5%——GoogleGemini 3.7 Flash model card2026-11-13 取得
2Gemini 3.6 Flash記載なし41.2%——GoogleGemini 3.7 Flash model card2026-11-13 取得

知識OfficeQA Pro v2

公式の結果(Introducing Kimi K3.1 Pro)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax64.4——Moonshot AIIntroducing Kimi K3.1 Pro2026-11-12 取得

知識Frontier Finance

公式の結果(Introducing Kimi K3.1 Pro)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax58.6——Moonshot AIIntroducing Kimi K3.1 Pro2026-11-12 取得

知識SimpleQA Verified

第三者の結果(Epoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.0)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2Max94.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
2Gemini 4 CarbonHigh77.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
3Gemini 4 ArgonHigh76.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
4GPT-6 AstraMax75.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
5GPT-6.1 SolMax73.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
6Gemini 3.1 Pro PreviewHigh73.5%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
7Claude Opus 5.5Max72.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
8Claude Fable 5Xhigh70.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
9Gemini 3.8 FlashHigh69.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
9GPT-5.6 SolMax69.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
11Gemini 3.7 FlashHigh69.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
12Gemini 3 Flash PreviewHigh66.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
13Gemini 3.6 FlashHigh66.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
14GPT-5.5Xhigh63.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
15Kimi K3.1 ProMax62.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
16GPT-6 SolMax60.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
17GPT-6.1 TerraMax57.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
18Grok 4.7Xhigh56.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
19Claude Opus 4.8Max53.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
20DeepSeek-V4-Pro-0813Max52.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
21Claude Opus 4.7Xhigh51.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
22Kimi K3Max50.6%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
23Grok 4.6High49.3%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
24Grok 4.6Xhigh48.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
25Claude Opus 4.6Max47.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
26Claude Sonnet 5.5Max46.5%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
27Qwen3.8-MaxXhigh45.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
28GPT-5.6 TerraMax43.2%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
29DeepSeek-V4.1-Flash記載なし42.9%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
30DeepSeek-V4.1-Flash-1017Max42.5%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
31GPT-6 LunaMax41.4%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
32GLM-5.3Max41.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
32GPT-5.6 LunaMax41.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
34Claude Sonnet 4.6High35.5%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
35Claude Sonnet 4.6Max32.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
36Claude Haiku 5.5Max23.8%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
37Claude Sonnet 4.5記載なし23.7%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得
38Mistral Large 4High20.0%——Epoch AIEpoch AI, Capabilities & benchmarking (benchmark_data.zip, simpleqa_verified.csv) CC BY 4.02026-11-12 取得

知識OfficeQA Pro

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax67.4——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max63.3——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

マルチモーダルChartography(ツールあり)

公式の結果(Introducing Claude Opus 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 5.5Max89.0%——AnthropicIntroducing Claude Opus 5.52026-11-12 取得
2Claude Fable 5.1記載なし88.4%——AnthropicIntroducing Claude Opus 5.52026-11-12 取得
3Claude Opus 5記載なし83.4%——AnthropicIntroducing Claude Opus 5.52026-11-12 取得

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax90.2%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(DeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1DeepSeek-V4.1-Flash記載なし78.9%——DeepSeekDeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai)2026-11-13 取得

マルチモーダルChartography(ツールなし)

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2Max73.4%$0.75Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Gemini 4 Carbon記載なし72.5%——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3Claude Fable 5.2Xhigh71.4%$0.15Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
4GPT-6 Astra記載なし71.0%——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
5Claude Fable 5.2High69.8%$0.07Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
6Claude Fable 5.2Med66.4%$0.03Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
7Claude Opus 5.5Max64.4%$0.49Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
8Claude Fable 5.2Low64.2%$0.03Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
9Claude Opus 5.5Xhigh62.4%$0.08Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
10Claude Opus 5.5High59.2%$0.03Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
11Claude Opus 5.5Med59.0%$0.03Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
12Claude Opus 5.5Low57.0%$0.02Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
13Claude Fable 5.1Xhigh47.0%$0.38Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
14Claude Fable 5.1High46.0%$0.08Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
15Claude Fable 5.1Max44.8%$0.87Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
16Claude Fable 5.1Med42.2%$0.05Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
17Claude Fable 5.1Low36.6%$0.03Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

公式の結果(Introducing Claude Sonnet 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 5.5記載なし64.4%——AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
2Claude Sonnet 5.5記載なし61.6%——AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
3GPT-6 Sol記載なし53.6%——記載なしIntroducing Claude Sonnet 5.52026-11-12 取得
4Claude Sonnet 5記載なし15.6%——AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得

公式の結果(Introducing Claude Haiku 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Sonnet 5.5記載なし61.6%——AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
2Claude Haiku 5.5記載なし46.4%——AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
3GPT-6 Luna記載なし29.1%——記載なしIntroducing Claude Haiku 5.52026-11-12 取得
4Claude Haiku 4.5記載なし6.4%——AnthropicIntroducing Claude Haiku 5.52026-11-12 取得

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh72.5%——記載なしGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh71.6%——記載なしGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

マルチモーダルBlueprint-Bench 2

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh58.6%——Andon LabsGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh54.4%——Andon LabsGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

マルチモーダルLVBench

公式の結果(Gemini 4 Argon model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 Argon記載なし91.7%——GoogleGemini 4 Argon model card2026-11-12 取得

公式の結果(Gemini 3.8 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.8 Flash記載なし87.8%——GoogleGemini 3.8 Flash model card2026-11-12 取得

公式の結果(Gemini 3.7 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.7 Flash記載なし85.4%——GoogleGemini 3.7 Flash model card2026-11-13 取得
2Gemini 3.6 Flash記載なし84.2%——GoogleGemini 3.7 Flash model card2026-11-13 取得

マルチモーダルCharXiv Reasoning

公式の結果(Gemini 3.8 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.8 Flash記載なし86.2%——GoogleGemini 3.8 Flash model card2026-11-12 取得

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax89.0%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max84.8%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Gemini 3.7 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.6 Flash記載なし85.2%——GoogleGemini 3.7 Flash model card2026-11-13 取得
2Gemini 3.7 Flash記載なし84.5%——GoogleGemini 3.7 Flash model card2026-11-13 取得

公式の結果(Gemini 3.5 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.5 Flash記載なし84.2%——GoogleGemini 3.5 Flash model card2026-11-13 取得
2Gemini 3.1 Pro Preview記載なし83.3%——GoogleGemini 3.5 Flash model card2026-11-13 取得
3Gemini 3 Flash Preview記載なし80.3%——GoogleGemini 3.5 Flash model card2026-11-13 取得

マルチモーダルCharXiv Reasoning(ツールあり)

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax93.4%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max91.3%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Gemini 3.7 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.6 Flash記載なし89.4%——GoogleGemini 3.7 Flash model card2026-11-13 取得
2Gemini 3.7 Flash記載なし88.7%——GoogleGemini 3.7 Flash model card2026-11-13 取得

マルチモーダルMMMU-Pro

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax84.1%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max81.6%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Gemini 3.5 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.5 Flash記載なし83.6%——GoogleGemini 3.5 Flash model card2026-11-13 取得
2Gemini 3 Flash Preview記載なし81.2%——GoogleGemini 3.5 Flash model card2026-11-13 取得
3Gemini 3.1 Pro Preview記載なし80.5%——GoogleGemini 3.5 Flash model card2026-11-13 取得

マルチモーダルMMMU-Pro(ツールあり)

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax90.0%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max83.4%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

マルチモーダルBabyVision(ツールあり)

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax95.4%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max85.7%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(DeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1DeepSeek-V4.1-Flash記載なし89.6%——DeepSeekDeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai)2026-11-13 取得

マルチモーダルZeroBench main(ツールあり・pass@5)

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax63.0%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max41.0%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(DeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1DeepSeek-V4.1-Flash記載なし49.0%——DeepSeekDeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai)2026-11-13 取得

長文脈GraphWalks BFS(256k〜1M)

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh96.2%——GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh84.2%——GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

長文脈GraphWalks BFS(128k 以下)

公式の結果(Gemini 4 Argon model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 Argon記載なし99.7%——GoogleGemini 4 Argon model card2026-11-12 取得

長文脈MRCR v2(8-needle・128k)

公式の結果(Gemini 3.7 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.7 Flash記載なし97.0%——GoogleGemini 3.7 Flash model card2026-11-13 取得
2Gemini 3.6 Flash記載なし91.8%——GoogleGemini 3.7 Flash model card2026-11-13 取得

公式の結果(Gemini 3.5 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.1 Pro Preview記載なし84.9%——GoogleGemini 3.5 Flash model card2026-11-13 取得
2Gemini 3.5 Flash記載なし77.3%——GoogleGemini 3.5 Flash model card2026-11-13 取得
3Gemini 3 Flash Preview記載なし67.2%——GoogleGemini 3.5 Flash model card2026-11-13 取得

長文脈MRCR v2(8-needle・1M)

公式の結果(Gemini 3.5 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.5 Flash記載なし26.6%——GoogleGemini 3.5 Flash model card2026-11-13 取得
2Gemini 3.1 Pro Preview記載なし26.3%——GoogleGemini 3.5 Flash model card2026-11-13 取得
3Gemini 3 Flash Preview記載なし22.1%——GoogleGemini 3.5 Flash model card2026-11-13 取得

ツール利用AutomationBench

公式の結果(Introducing Claude Opus 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GPT-6 AstraMax41.4%$1.77Cost per task前線ZapierIntroducing Claude Opus 5.52026-11-12 取得
2Claude Opus 5.5Max40.0%$1.37Cost per task前線ZapierIntroducing Claude Opus 5.52026-11-12 取得
3GPT-6 AstraXhigh39.0%$1.53Cost per task—ZapierIntroducing Claude Opus 5.52026-11-12 取得
4GPT-6 AstraHigh37.1%$1.45Cost per task—ZapierIntroducing Claude Opus 5.52026-11-12 取得
5Claude Opus 5.5Xhigh34.4%$0.86Cost per task前線ZapierIntroducing Claude Opus 5.52026-11-12 取得
6GPT-6 AstraMed34.1%$1.28Cost per task—ZapierIntroducing Claude Opus 5.52026-11-12 取得
7Claude Opus 5.5High32.0%$0.70Cost per task前線ZapierIntroducing Claude Opus 5.52026-11-12 取得
8Claude Fable 5.1記載なし31.4%——ZapierIntroducing Claude Opus 5.52026-11-12 取得
9GPT-6 AstraLow30.3%$1.08Cost per task—ZapierIntroducing Claude Opus 5.52026-11-12 取得
10GPT-5.6 SolMax28.8%$0.91Cost per task—ZapierIntroducing Claude Opus 5.52026-11-12 取得
11Claude Opus 5.5Med28.6%$0.64Cost per task前線ZapierIntroducing Claude Opus 5.52026-11-12 取得
12Claude Opus 5Max26.9%$1.27Cost per task—ZapierIntroducing Claude Opus 5.52026-11-12 取得
13GPT-5.6 SolXhigh26.3%$0.74Cost per task—ZapierIntroducing Claude Opus 5.52026-11-12 取得
14Claude Opus 5Xhigh25.3%$1.15Cost per task—ZapierIntroducing Claude Opus 5.52026-11-12 取得
15GPT-5.6 SolHigh24.8%$0.64Cost per task—ZapierIntroducing Claude Opus 5.52026-11-12 取得
16Claude Opus 5Med23.9%$0.89Cost per task—ZapierIntroducing Claude Opus 5.52026-11-12 取得
17Claude Opus 5.5Low23.3%$0.50Cost per task前線ZapierIntroducing Claude Opus 5.52026-11-12 取得
18Claude Opus 5High20.6%$1.03Cost per task—ZapierIntroducing Claude Opus 5.52026-11-12 取得
19Claude Opus 5Low20.4%$0.75Cost per task—ZapierIntroducing Claude Opus 5.52026-11-12 取得
20GPT-5.6 SolMed19.6%$0.57Cost per task—ZapierIntroducing Claude Opus 5.52026-11-12 取得
21GPT-5.6 SolLow11.7%$0.42Cost per task前線ZapierIntroducing Claude Opus 5.52026-11-12 取得

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 Carbon記載なし53.9%——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Claude Fable 5.2記載なし51.4%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3Claude Opus 5.5記載なし42.5%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
4GPT-6 Astra記載なし41.4%——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
5Claude Fable 5.1記載なし31.4%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh53.9%——ZapierGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh51.3%——ZapierGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

公式の結果(Introducing Mistral Large 4)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Mistral Large 4記載なし59.9%——Mistral AIIntroducing Mistral Large 42026-11-12 取得

ツール利用WANDR

公式の結果(Introducing Claude Opus 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 5.5Max72.3%$37.92Cost per attempt前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
2Claude Opus 5.5Xhigh71.3%$29.06Cost per attempt前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
3Claude Fable 5.1Max68.7%$49.02Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
4Claude Fable 5.1Xhigh67.7%$42.81Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
5Claude Opus 5.5High67.3%$16.88Cost per attempt前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
6Claude Opus 5Max67.2%$61.62Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
7Claude Opus 5Xhigh67.0%$53.84Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
8Claude Fable 5.1High66.7%$33.62Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
9Claude Opus 5High64.6%$43.49Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
10Claude Fable 5.1Med64.5%$27.41Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
11Claude Fable 5.1Low63.3%$23.25Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
12Claude Opus 5.5Med62.8%$11.20Cost per attempt前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
13Claude Opus 5Med58.1%$23.97Cost per attempt—AnthropicIntroducing Claude Opus 5.52026-11-12 取得
14Claude Opus 5Low50.5%$10.65Cost per attempt前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得
15Claude Opus 5.5Low31.2%$1.20Cost per attempt前線AnthropicIntroducing Claude Opus 5.52026-11-12 取得

ツール利用WANDR(relaxed score)

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2Max82.1%$53.11Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Claude Fable 5.2Xhigh80.6%$39.84Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3Claude Fable 5.2High79.3%$29.20Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
4Claude Fable 5.2Med73.6%$15.71Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
5Claude Opus 5.5Max72.3%$38.68Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
6Claude Fable 5.2Low71.4%$12.66Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
7Claude Opus 5.5Xhigh71.3%$29.25Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
8Claude Fable 5.1Max68.7%$51.10Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
9Claude Fable 5.1Xhigh67.7%$44.50Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
10Claude Opus 5.5High67.3%$17.31Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
11Claude Fable 5.1High66.7%$37.60Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
12Claude Fable 5.1Med64.5%$28.78Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
13Claude Fable 5.1Low63.3%$24.41Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
14Claude Opus 5.5Med62.8%$11.40Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
15Claude Opus 5.5Low31.2%$1.25Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

ツール利用Vals Finance Agent v2

公式の結果(Gemini 4 Argon model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 Argon記載なし65.4%——GoogleGemini 4 Argon model card2026-11-12 取得

公式の結果(Gemini 3.8 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.8 Flash記載なし61.4%——GoogleGemini 3.8 Flash model card2026-11-12 取得

ツール利用AutomationBench v1.0.6

公式の結果(GLM-5.3 (Hugging Face model card, zai-org))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GLM-5.3記載なし48.2%——Z.aiGLM-5.3 (Hugging Face model card, zai-org)2026-11-13 取得

公式の結果(GLM-5.3-Flash (Z.ai docs))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GLM-5.3-Flash記載なし48.8%——Z.aiGLM-5.3-Flash (Z.ai docs)2026-11-13 取得

公式の結果(DeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1DeepSeek-V4.1-Flash記載なし54.8%——DeepSeekDeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai)2026-11-13 取得

ツール利用AutomationBench(公開 600 タスク)

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max30.8%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Qwen3.8-2.4T-A95B (Hugging Face model card, Qwen))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Qwen3.8-Max記載なし27.3%——Alibaba (Qwen)Qwen3.8-2.4T-A95B (Hugging Face model card, Qwen)2026-11-13 取得

ツール利用BrowseComp

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax92.1%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max91.2%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

ツール利用MCP Atlas

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max84.2%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

ツール利用τ²-bench airline (v1.0.1)

第三者の結果(τ²-bench v1.0.1(Sierra の実行))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 4.5High84.0%$0.40Average cost in USD to run one trajectory in this domain前線Sierraτ²-bench leaderboard submission claude-opus-4-5_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得
2GPT-5.2High83.0%$0.11Average cost in USD to run one trajectory in this domain前線Sierraτ²-bench leaderboard submission gpt-5-2_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得
3Gemini 3 Flash PreviewHigh82.5%——Sierraτ²-bench leaderboard submission gemini-3-flash_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得
3GLM-5記載なし82.5%——Sierraτ²-bench leaderboard submission glm-5-think_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得
5Qwen3.5-397B-A17B記載なし81.5%——Sierraτ²-bench leaderboard submission qwen3.5-397b-a17b-think_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得
6Gemini 3 Pro PreviewHigh80.5%——Sierraτ²-bench leaderboard submission gemini-3-pro_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得
7Claude Sonnet 4.5記載なし72.0%$0.30Average cost in USD to run one trajectory in this domain—Sierraτ²-bench leaderboard submission claude-sonnet-4-5_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得
8GPT-5.2None52.5%$0.054Average cost in USD to run one trajectory in this domain前線Sierraτ²-bench leaderboard submission gpt-5-2-none_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得

ツール利用τ²-bench banking_knowledge (v1.0.1)

第三者の結果(τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 alltools))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Qwen3.8-MaxXhigh55.2%——Sierraτ²-bench leaderboard submission qwen3-8-max_sierra_2026-08-04 (sierra-research/tau2-bench, MIT)2026-11-13 取得
2Claude Opus 5Max48.7%——Sierraτ²-bench leaderboard submission claude-opus-5_sierra_2026-08-04 (sierra-research/tau2-bench, MIT)2026-11-13 取得
3Grok 4.5High47.9%——Sierraτ²-bench leaderboard submission grok-4-5_sierra_2026-08-04 (sierra-research/tau2-bench, MIT)2026-11-13 取得
4GPT-5.6 SolXhigh46.9%——Sierraτ²-bench leaderboard submission gpt-5-6-sol_sierra_2026-08-04 (sierra-research/tau2-bench, MIT)2026-11-13 取得
5GPT-5.5Xhigh44.6%——Sierraτ²-bench leaderboard submission gpt-5-5_sierra_2026-05-05 (sierra-research/tau2-bench, MIT)2026-11-13 取得
6Claude Opus 4.7Max40.2%——Sierraτ²-bench leaderboard submission claude-opus-4-7_sierra_2026-05-05 (sierra-research/tau2-bench, MIT)2026-11-13 取得
7Claude Fable 5Max39.7%——Sierraτ²-bench leaderboard submission claude-fable-5_sierra_2026-08-04 (sierra-research/tau2-bench, MIT)2026-11-13 取得
7Claude Opus 4.8Max39.7%——Sierraτ²-bench leaderboard submission claude-opus-4-8_sierra_2026-08-04 (sierra-research/tau2-bench, MIT)2026-11-13 取得
9GPT-5.4Xhigh39.4%$1.84Average cost in USD to run one trajectory in this domain前線Sierraτ²-bench leaderboard submission gpt-5-4_sierra_2026-03-25 (sierra-research/tau2-bench, MIT)2026-11-13 取得
10GLM-5.2記載なし37.1%——Sierraτ²-bench leaderboard submission glm-5-2_sierra_2026-08-04 (sierra-research/tau2-bench, MIT)2026-11-13 取得
10Kimi K3Max37.1%——Sierraτ²-bench leaderboard submission kimi-k3_sierra_2026-08-04 (sierra-research/tau2-bench, MIT)2026-11-13 取得
12GPT-5.2High32.2%——Sierraτ²-bench leaderboard submission gpt-5-2_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得
13Claude Opus 4.6Max27.3%——Sierraτ²-bench leaderboard submission claude-opus-4-6_sierra_2026-05-05 (sierra-research/tau2-bench, MIT)2026-11-13 取得
14Gemini 3.1 Pro PreviewHigh26.0%——Sierraτ²-bench leaderboard submission gemini-3-1-pro-preview_sierra_2026-05-05 (sierra-research/tau2-bench, MIT)2026-11-13 取得
15Claude Opus 4.5High24.7%——Sierraτ²-bench leaderboard submission claude-opus-4-5_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得
16Gemini 2.5 ProHigh13.7%——Sierraτ²-bench leaderboard submission gemini-2-5-pro_sierra_2026-05-05 (sierra-research/tau2-bench, MIT)2026-11-13 取得

第三者の結果(τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 terminal))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3 Flash PreviewHigh27.3%——Sierraτ²-bench leaderboard submission gemini-3-flash_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得
2Claude Sonnet 4.5記載なし25.3%$4.05Average cost in USD to run one trajectory in this domain前線Sierraτ²-bench leaderboard submission claude-sonnet-4-5_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得
3Gemini 3 Pro PreviewHigh18.0%——Sierraτ²-bench leaderboard submission gemini-3-pro_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得

第三者の結果(τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 text-emb-3-large))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GLM-5記載なし9.8%——Sierraτ²-bench leaderboard submission glm-5-think_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得
1Qwen3.5-397B-A17B記載なし9.8%——Sierraτ²-bench leaderboard submission qwen3.5-397b-a17b-think_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得

第三者の結果(τ²-bench v1.0.1 banking_knowledge(Sierra の実行・検索 qwen_embeddings))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GPT-5.2None12.6%$0.24Average cost in USD to run one trajectory in this domain前線Sierraτ²-bench leaderboard submission gpt-5-2-none_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得

ツール利用τ²-bench retail (v1.0.1)

第三者の結果(τ²-bench v1.0.1(Sierra の実行))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Qwen3.5-397B-A17B記載なし84.4%——Sierraτ²-bench leaderboard submission qwen3.5-397b-a17b-think_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得
2GPT-5.2High81.6%$0.072Average cost in USD to run one trajectory in this domain前線Sierraτ²-bench leaderboard submission gpt-5-2_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得
3Claude Opus 4.5High79.6%$0.39Average cost in USD to run one trajectory in this domain—Sierraτ²-bench leaderboard submission claude-opus-4-5_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得
4Gemini 3 Flash PreviewHigh76.8%——Sierraτ²-bench leaderboard submission gemini-3-flash_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得
5Gemini 3 Pro PreviewHigh75.9%——Sierraτ²-bench leaderboard submission gemini-3-pro_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得
6GPT-5.2None75.0%$0.048Average cost in USD to run one trajectory in this domain前線Sierraτ²-bench leaderboard submission gpt-5-2-none_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得
7GLM-5記載なし73.7%——Sierraτ²-bench leaderboard submission glm-5-think_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得
8Claude Sonnet 4.5記載なし72.4%$0.25Average cost in USD to run one trajectory in this domain—Sierraτ²-bench leaderboard submission claude-sonnet-4-5_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得

ツール利用τ²-bench telecom (v1.0.1)

第三者の結果(τ²-bench v1.0.1(Sierra の実行))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Qwen3.5-397B-A17B記載なし97.8%——Sierraτ²-bench leaderboard submission qwen3.5-397b-a17b-think_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得
2Claude Opus 4.5High92.3%$0.72Average cost in USD to run one trajectory in this domain前線Sierraτ²-bench leaderboard submission claude-opus-4-5_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得
3Gemini 3 Flash PreviewHigh91.2%——Sierraτ²-bench leaderboard submission gemini-3-flash_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得
4Gemini 3 Pro PreviewHigh91.0%——Sierraτ²-bench leaderboard submission gemini-3-pro_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得
5GPT-5.2High89.7%$0.093Average cost in USD to run one trajectory in this domain前線Sierraτ²-bench leaderboard submission gpt-5-2_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得
6GLM-5記載なし86.8%——Sierraτ²-bench leaderboard submission glm-5-think_sierra_2026-03-02 (sierra-research/tau2-bench, MIT)2026-11-13 取得
7Claude Sonnet 4.5記載なし84.9%$0.49Average cost in USD to run one trajectory in this domain—Sierraτ²-bench leaderboard submission claude-sonnet-4-5_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得
8GPT-5.2None57.2%$0.049Average cost in USD to run one trajectory in this domain前線Sierraτ²-bench leaderboard submission gpt-5-2-none_sierra_2026-02-26 (sierra-research/tau2-bench, MIT)2026-11-13 取得

コンピュータ操作OSWorld 2.1(partial)

公式の結果(Introducing Claude Opus 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 5.5Max81.8%——AnthropicIntroducing Claude Opus 5.52026-11-12 取得
2Claude Fable 5.1記載なし80.7%——AnthropicIntroducing Claude Opus 5.52026-11-12 取得
3Claude Opus 5記載なし74.0%——AnthropicIntroducing Claude Opus 5.52026-11-12 取得

公式の結果(Introducing Claude Sonnet 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 5.5記載なし81.8%——AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
2Claude Sonnet 5.5記載なし80.1%——AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得
3Claude Sonnet 5記載なし57.0%——AnthropicIntroducing Claude Sonnet 5.52026-11-12 取得

コンピュータ操作OSWorld 2.1(offline subset)

公式の結果(Introducing Claude Haiku 5.5)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Sonnet 5.5Max83.9%$5.73Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
2Claude Sonnet 5.5Xhigh81.1%$2.22Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
3Claude Sonnet 5.5High73.2%$1.38Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
4Claude Haiku 5.5Max72.4%$0.61Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
5Claude Haiku 5.5Xhigh67.6%$0.28Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
6Claude Sonnet 5.5Med66.0%$0.93Cost per attempt—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
7Claude Haiku 5.5High61.3%$0.18Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
8Claude Sonnet 5.5Low57.9%$0.68Cost per attempt—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
9Claude Haiku 5.5Med53.3%$0.13Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
10GPT-6 LunaMax48.9%$0.20Cost per attempt—記載なしIntroducing Claude Haiku 5.52026-11-12 取得
11GPT-6 LunaXhigh44.8%$0.17Cost per attempt—記載なしIntroducing Claude Haiku 5.52026-11-12 取得
12GPT-6 LunaHigh42.3%$0.14Cost per attempt—記載なしIntroducing Claude Haiku 5.52026-11-12 取得
13Claude Haiku 5.5Low42.0%$0.07Cost per attempt前線AnthropicIntroducing Claude Haiku 5.52026-11-12 取得
14GPT-6 LunaMed37.5%$0.13Cost per attempt—記載なしIntroducing Claude Haiku 5.52026-11-12 取得
15GPT-6 LunaLow19.2%$0.04Cost per attempt前線記載なしIntroducing Claude Haiku 5.52026-11-12 取得
16Claude Haiku 4.5記載なし15.7%$1.45Cost per attempt—AnthropicIntroducing Claude Haiku 5.52026-11-12 取得

コンピュータ操作OSWorld 2.1(offline set・partial)

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2記載なし89.3%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Claude Opus 5.5記載なし87.2%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3Claude Fable 5.1記載なし83.5%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
4Gemini 4 Carbon記載なし81.5%——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
5GPT-6 Astra記載なし79.5%——記載なしIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh81.5%——GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh74.7%——GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

コンピュータ操作OSWorld 2.0

公式の結果(Gemini 4 Argon model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 Argon記載なし69.2%——GoogleGemini 4 Argon model card2026-11-12 取得

公式の結果(Gemini 3.8 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.8 Flash記載なし59.0%——GoogleGemini 3.8 Flash model card2026-11-12 取得

公式の結果(Gemini 3.7 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.7 Flash記載なし47.9%——GoogleGemini 3.7 Flash model card2026-11-13 取得
2Gemini 3.6 Flash記載なし33.8%——GoogleGemini 3.7 Flash model card2026-11-13 取得

コンピュータ操作Agents' Last Exam v1.1(rubric score)

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh77.4%——GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh69.0%——GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

コンピュータ操作Agents' Last Exam(pass rate・版の記載なし)

公式の結果(Gemini 4 Argon model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 Argon記載なし39.5%——GoogleGemini 4 Argon model card2026-11-12 取得

公式の結果(Gemini 3.7 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.7 Flash記載なし26.3%——GoogleGemini 3.7 Flash model card2026-11-13 取得
2Gemini 3.6 Flash記載なし24.2%——GoogleGemini 3.7 Flash model card2026-11-13 取得

公式の結果(DeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1DeepSeek-V4.1-Flash記載なし31.8%——DeepSeekDeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai)2026-11-13 取得

コンピュータ操作Agents' Last Exam(版・指標の記載なし)

公式の結果(GLM-5.3 (Z.ai docs))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GLM-5.3記載なし28.5%——Z.aiGLM-5.3 (Z.ai docs)2026-11-12 取得

コンピュータ操作OSWorld-Verified

公式の結果(Gemini 3.5 Flash model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 3.5 Flash記載なし78.4%——GoogleGemini 3.5 Flash model card2026-11-13 取得
2Gemini 3.1 Pro Preview記載なし76.2%——GoogleGemini 3.5 Flash model card2026-11-13 取得
3Gemini 3 Flash Preview記載なし65.1%——GoogleGemini 3.5 Flash model card2026-11-13 取得

公式の結果(Introducing Claude Opus 4.8)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 4.7記載なし82.3%——AnthropicIntroducing Claude Opus 4.82026-11-13 取得

その他BenchCAD 2.1(ツールあり)

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2Max83.5$50.77Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Claude Fable 5.2Xhigh74.1$23.23Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3Claude Opus 5.5Max65.3$19.86Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
4Claude Fable 5.2High61$9.99Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
5Claude Opus 5.5Xhigh54.5$8.91Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
6Claude Fable 5.1Max44.4$34.69Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
7Claude Fable 5.1Xhigh40.5$27.38Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
8Claude Fable 5.2Med32.4$2.66Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
9Claude Opus 5.5High29.7$2.57Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
10Claude Fable 5.2Low26.9$1.89Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
11Claude Fable 5.1High25.8$12.44Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
12Claude Opus 5.5Med19.1$1.11Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
13Claude Fable 5.1Med15.7$5.70Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
14Claude Fable 5.1Low12.1$2.95Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
15Claude Opus 5.5Low6.3$0.03Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

その他BenchCAD(ツールあり・版の記載なし)

公式の結果(Introducing Kimi K3.1 Pro)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax92.7——Moonshot AIIntroducing Kimi K3.1 Pro2026-11-12 取得

その他BenchCAD(ツールなし・版の記載なし)

公式の結果(Introducing Kimi K3.1 Pro)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax72.8——Moonshot AIIntroducing Kimi K3.1 Pro2026-11-12 取得

その他CVE Rediscovery

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Mythos 5.2Max89.6%$9.13Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Claude Mythos 5.2Xhigh88.1%$2.40Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3Claude Mythos 5.2High83.2%$1.01Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
4Claude Opus 5.5Max81.3%$2.98Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
5Claude Mythos 5.1Max78.3%$4.14Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
6Claude Mythos 5.1Xhigh77.8%$2.81Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
7Claude Opus 5.5Xhigh76.4%$1.05Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
8Claude Mythos 5.1High73.8%$1.39Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
9Claude Mythos 5.2Med73.3%$0.40Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
10Claude Mythos 5.2Low72.4%$0.34Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
11Claude Opus 5.5High70.0%$0.43Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
12Claude Mythos 5.1Med69.8%$0.89Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
13Claude Opus 5.5Med66.1%$0.28Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
14Claude Mythos 5.1Low65.9%$0.59Cost per attempt—AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
15Claude Opus 5.5Low59.4%$0.11Cost per attempt前線AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

その他CyScenarioBench

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Mythos 5.2記載なし76.3%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Claude Opus 5.5記載なし67.6%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3Claude Mythos 5.1記載なし61.7%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

その他De novo binder design(24 時間)

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Mythos 5.2記載なし85.1——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Claude Opus 5.5記載なし82.6——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3Claude Mythos 5.1記載なし79.1——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

その他RoboDojo Sim

公式の結果(Introducing Claude Fable 5.2 and Claude Mythos 5.2)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2記載なし64.9%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
2Claude Opus 5.5記載なし39.4%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
3GPT-6 Astra記載なし29.8%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得
4Claude Fable 5.1記載なし27.6%——AnthropicIntroducing Claude Fable 5.2 and Claude Mythos 5.22026-11-12 取得

その他Artificial Analysis Intelligence Index v5.0

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh59——Artificial AnalysisGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh50——Artificial AnalysisGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

その他Vals Index v2.2

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh70.4%——Vals AIGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh69.2%——Vals AIGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

その他Vending-Bench 2

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh16,484——Andon LabsGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh13,718——Andon LabsGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

その他ExploitGym

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh54.8%——GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh42.0%——GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax20.7%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得

その他SEC-bench Pro

公式の結果(Gemini 4 Carbon (Google DeepMind models page, benchmark table))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 CarbonHigh88.2%——GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得
2Gemini 4 ArgonHigh83.4%——GoogleGemini 4 Carbon (Google DeepMind models page, benchmark table)2026-11-12 取得

公式の結果(DeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1DeepSeek-V4.1-Flash記載なし62.8%——DeepSeekDeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai)2026-11-13 取得

その他PostTrainBench

公式の結果(Gemini 4 Argon model card)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Gemini 4 Argon記載なし45.3%——GoogleGemini 4 Argon model card2026-11-12 取得

公式の結果(Kimi K3 (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3Max36.6%——Moonshot AIKimi K3 (Hugging Face model card, moonshotai)2026-11-13 取得

公式の結果(GLM-5.3 (Hugging Face model card, zai-org))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GLM-5.3Max39.8%——Z.aiGLM-5.3 (Hugging Face model card, zai-org)2026-11-13 取得

その他EEBench

公式の結果(Grok 4.7 (xAI news))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Grok 4.7Xhigh64.0%——xAIGrok 4.7 (xAI news)2026-11-12 取得
2Grok 4.6High53.0%——xAIGrok 4.7 (xAI news)2026-11-12 取得

その他CyberGym

公式の結果(GLM-5.3 (Z.ai docs))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GLM-5.3記載なし84.5%——Z.aiGLM-5.3 (Z.ai docs)2026-11-12 取得

公式の結果(Introducing Claude Opus 4.7)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Opus 4.6記載なし73.8%——AnthropicIntroducing Claude Opus 4.72026-11-13 取得

その他ExploitBench

公式の結果(GLM-5.3 (Z.ai docs))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1GLM-5.3記載なし54.4%——Z.aiGLM-5.3 (Z.ai docs)2026-11-12 取得

その他Cybench

公式の結果(Introducing Mistral Large 4)

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Mistral Large 4記載なし93.0%——Mistral AIIntroducing Mistral Large 42026-11-12 取得

その他Epoch Capabilities Index(ECI)

第三者の結果(Epoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Claude Fable 5.2記載なし173.44——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
2Gemini 4 Carbon記載なし169.76——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
3GPT-6 Astra記載なし169——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
4Claude Opus 5.5記載なし168.54——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
5GPT-6.1 Sol記載なし168.48——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
6GPT-6.1 Terra記載なし167.67——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
7Claude Sonnet 5.5記載なし166.56——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
8Gemini 4 Argon記載なし166.48——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
9Claude Fable 5記載なし164.13——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
10GPT-6 Sol記載なし163.75——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
11GPT-5.6 Sol記載なし162.99——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
12Kimi K3.1 Pro記載なし161.37——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
13GPT-5.6 Terra記載なし160.73——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
14Claude Haiku 5.5記載なし160.43——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
15GPT-5.5記載なし160.1——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
16Claude Opus 4.8記載なし159.94——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
17DeepSeek-V4.1-Flash-1017記載なし159.41——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
18Gemini 3.8 Flash記載なし158.79——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
19Kimi K3記載なし158.58——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
20Gemini 3.7 Flash記載なし158.36——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
21Grok 4.6記載なし157.79——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
22Qwen3.8-Max記載なし157.59——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
23Claude Opus 4.7記載なし157.44——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
24GLM-5.3記載なし157.1——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
25DeepSeek-V4.1-Flash記載なし156.7——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
26GPT-5.6 Luna記載なし156.67——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
27GPT-6 Luna記載なし156.41——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
28Claude Opus 4.6記載なし156.34——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
29Gemini 3.1 Pro Preview記載なし156.18——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
29Grok 4.7記載なし156.18——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
31DeepSeek-V4-Pro-0813記載なし155.94——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
32Gemini 3.6 Flash記載なし155.21——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
33GLM-5.3-Flash記載なし153.82——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
34Claude Sonnet 4.6記載なし152.86——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
35Mistral Large 4記載なし152.69——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
36Gemini 3 Flash Preview記載なし152.32——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
37Claude Opus 4.5記載なし151.92——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
38Claude Sonnet 4.5記載なし147.63——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
39Gemini 3.5 Flash-Lite記載なし145.63——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
40Gemini 3.1 Flash-Lite記載なし145.01——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
41Gemini 2.5 Pro記載なし144.87——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得
42Gemini 2.5 Flash記載なし140.81——Epoch AIEpoch AI, epoch_capabilities_index/eci_scores.csv (CC BY 4.0)2026-11-12 取得

その他PostTrainBench v1.2

公式の結果(Kimi K3.1 Pro (Hugging Face model card, moonshotai))

順位モデルeffortスコア出典のコスト(指標)前線計測・報告出典
1Kimi K3.1 ProMax36.5%——Moonshot AIKimi K3.1 Pro (Hugging Face model card, moonshotai)2026-11-13 取得