| Benchmark | Claude 3.5 Sonnet | o3 |
|---|---|---|
| AA Intelligence | 10 | 31.1 |
| aider_polyglot | 45.3 | 81.3 |
| AIME 2024 | 16 | 91.6 |
| AIME 2025 | 7.4 | 89.2 |
| AIR-Bench 2024 | 85.9 | 84.5 |
| anthropic_red_team | 99.8 | 98.3 |
| arena_vision | 1146 | 1217 |
| Artificial Analysis Coding Index | 30.2 | 38.4 |
| bbq | 94.9 | 97.9 |
| Fortress | 13 | 16 |
| frontiermath_tier_4 | 0 | 2.1 |
| GPQA Diamond | 67.2 | 83.3 |
| harmbench | 98.1 | 98.4 |
| HLE | 3.9 | 20.3 |
| hmmt_feb_2025 | 1.7 | 77.5 |
| LiveCodeBench | 32.8 | 84.7 |
| longbench_v2 | 41 | 58.8 |
| math_500_em | 78.3 | 98.1 |
| MathVista | 67.7 | 86.8 |
| MMLU-Pro | 78 | 85 |
| MMMU | 72 | 82.9 |
| MMMU-Pro | 54.7 | 76.4 |
| scicode | 36.6 | 41 |
| simple_safety_tests | 100 | 99 |
| simplebench | 41.4 | 53.1 |
| simpleqa | 28.4 | 49.4 |
| SWE-bench Verified | 50.8 | 69.1 |
| TAU-bench (airline) | 46 | 52 |
| TAU-bench (retail) | 69.2 | 73.9 |
| xstest | 95.6 | 97.3 |
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.