| Benchmark | o3 | Qwen2.5 Instruct 72B |
|---|---|---|
| AA Intelligence | 31.1 | 10 |
| AA-LCR | 73.3 | 20.3 |
| AA-Omniscience | -15.3 | -52.2 |
| aider_polyglot | 81.3 | 7.6 |
| AIME 2024 | 91.6 | 23.3 |
| AIR-Bench 2024 | 84.5 | 59 |
| anthropic_red_team | 98.3 | 99.6 |
| Artificial Analysis Coding Index | 38.4 | 11.9 |
| bbq | 97.9 | 95.4 |
| critpt | 1.1 | 0 |
| Fortress | 16 | 56.4 |
| GPQA Diamond | 83.3 | 49.1 |
| harmbench | 98.4 | 72.8 |
| HLE | 20.3 | 4.2 |
| IFBench | 71.4 | 36.9 |
| LiveCodeBench | 84.7 | 55.5 |
| longbench_v2 | 58.8 | 39.4 |
| math_500_em | 98.1 | 80 |
| MMLU-Pro | 85 | 71.6 |
| OmniScience Accuracy | 38.6 | 17.6 |
| OmniScience Non-Hallucination | 12.9 | 15.3 |
| scicode | 41 | 26.7 |
| simple_safety_tests | 99 | 100 |
| simpleqa | 49.4 | 10.3 |
| SWE-bench Verified | 69.1 | 23.8 |
| Terminal-Bench Hard | 37.1 | 4.5 |
| xstest | 97.3 | 97.9 |
| τ²-Bench Telecom (AA run) | 80.7 | 34.5 |
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.