| Benchmark | Claude 3.5 Sonnet | Gemini 2.5 Pro |
|---|---|---|
| AA Intelligence | 10 | 27 |
| aider_polyglot | 45.3 | 82.2 |
| AIME 2025 | 7.4 | 88.3 |
| AIR-Bench 2024 | 85.9 | 73.6 |
| anthropic_red_team | 99.8 | 99.5 |
| arena_vision | 1146 | 1246 |
| Artificial Analysis Coding Index | 30.2 | 46.7 |
| bbq | 94.9 | 96.4 |
| Fortress | 13 | 54.9 |
| frontiermath_tier_4 | 0 | 4.2 |
| GPQA Diamond | 67.2 | 84.4 |
| harmbench | 98.1 | 65.4 |
| HLE | 3.9 | 22.5 |
| hmmt_feb_2025 | 1.7 | 82.5 |
| LiveCodeBench | 32.8 | 82.7 |
| MathVista | 67.7 | 83.9 |
| MMLU-Pro | 78 | 86 |
| MMMU | 72 | 83.9 |
| MMMU-Pro | 54.7 | 74.9 |
| MMStar | 62.2 | 79.2 |
| OCRBench | 790 | 85.9 |
| scicode | 36.6 | 43 |
| SEAL VISTA | 38.7 | 50.8 |
| simple_safety_tests | 100 | 97 |
| simplebench | 41.4 | 62.4 |
| simpleqa | 28.4 | 50.8 |
| SWE-bench Verified | 50.8 | 67.2 |
| xstest | 95.6 | 98.7 |
Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing.