| Benchmark | Claude 3.5 Sonnet | GPT-5 |
|---|---|---|
| AA Intelligence | 10 | 35.3 |
| aider_polyglot | 45.3 | 88 |
| AIME 2025 | 7.4 | 95 |
| AIR-Bench 2024 | 85.9 | 87.7 |
| arena_vision | 1146 | 1212 |
| Artificial Analysis Coding Index | 30.2 | 38.9 |
| Fortress | 13 | 17 |
| frontiermath_tier_4 | 0 | 12.5 |
| GPQA Diamond | 67.2 | 87.3 |
| GSM8K | 96.9 | -1.8 |
| harmbench | 98.1 | 70 |
| HLE | 3.9 | 28.5 |
| hmmt_feb_2025 | 1.7 | 88.3 |
| humaneval | 93.7 | 93.4 |
| LiveCodeBench | 32.8 | 85 |
| LiveCodeBench v6 | 37.2 | 87 |
| mmlu_redux | 88.9 | 95.3 |
| MMLU-Pro | 78 | 87.5 |
| MMMU | 72 | 84.2 |
| MMMU-Pro | 54.7 | 78.4 |
| scicode | 36.6 | 43 |
| SEAL VISTA | 38.7 | 49.7 |
| simplebench | 41.4 | 56.7 |
| SWE-bench Verified | 50.8 | 74.9 |
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.