| Benchmark | Gemini 1.5 Pro | o3 |
|---|---|---|
| AA Intelligence | 10 | 31.1 |
| aider_polyglot | 16.9 | 81.3 |
| AIR-Bench 2024 | 82.8 | 84.5 |
| anthropic_red_team | 99.9 | 98.3 |
| ARC-AGI-2 | 0.8 | 6.5 |
| arena_vision | 1179 | 1217 |
| Artificial Analysis Coding Index | 23.6 | 38.4 |
| bbq | 94.5 | 97.9 |
| Fortress | 53.9 | 16 |
| GPQA Diamond | 59.1 | 83.3 |
| harmbench | 79.9 | 98.4 |
| HLE | 4.9 | 20.3 |
| LiveCodeBench | 30.5 | 84.7 |
| MathVista | 70.6 | 86.8 |
| MMLU-Pro | 75.8 | 85 |
| MMMU | 68.4 | 82.9 |
| MMMU-Pro | 55 | 76.4 |
| scicode | 29.5 | 41 |
| simple_safety_tests | 97.5 | 99 |
| simplebench | 27.1 | 53.1 |
| simpleqa | 23.4 | 49.4 |
| SWE-bench Verified | 34.2 | 69.1 |
| xstest | 98.8 | 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.