| SWE-Bench Verified | ~83% (est) | 87.6% | 85.0% | 82.1% | 81.0% | 80.6% |
| Terminal-Bench 2.0/2.1 | 74.6% | 66.1–69.4% | 78.2–82.7% | 76.2% | 68.5% | 65.0% |
| OSWorld-Verified(コンピュータ操作) | 83.4% | 82.8% | 78.7% | 75.0% | 72.1% | 70.5% |
| GDPval-AA(知識労働 / エージェントELO) | 1890 | 1753 | 1620–1769 | 1656 | 1500–1570 | 1550 |
| Humanity's Last Exam(ツールあり) | 57.9% | 54.7% | — | — | — | — |
| Humanity's Last Exam(ツールなし) | 49.8% | — | — | — | — | — |
| GPQA Diamond(博士課程レベルの科学推論) | ~94% (est) | 94.2% | 96.0% | 92.4% | 90.1–91.5% | 95.1% |
| ARC-AGI-2(抽象的視覚推論) | ~80% (est) | 80.2% | 85.0% | 75.8% | 76.1% | 74.0% |
| MCP Atlas(ツール使用の信頼性) | — | 77.3% | 79.1% | 83.6% | 74.2% | 71.5% |
| AA Intelligence Index (v4.0) | ~59–60 (est) | 59 | 60 | 57 | 53 | 55 |
| Finance Agent v2 | 53.9% | 51.5% | — | — | — | — |
| LiveCodeBench (Pass@1) | — | — | ~91–92% (est) | — | — | 93.5% |
| Codeforces ELO | — | ~3050 (est) | 3168 | — | — | 3206 |
| FrontierMath Tier 1–3 | — | 43.8% | 51.7% | — | — | — |
| MMLU-Pro | — | — | — | — | — | 87.5% |
| AIME 2025(数学) | — | — | 95.2% | — | — | — |
| BrowseComp | — | 79.3% | 84.4% | — | — | — |
* Gemini 3.5 Flash、Grok 4.3、DeepSeek V4 ProのSWE-Bench Proスコアは、単一の第三者テスト(出典)によるものです。Googleの公式モデルカードでは異なる数値が示されています。
Comments
0 comments