| Benchmark | Gemini 2.5 Pro | Qwen2.5 Instruct 72B |
|---|---|---|
| AA Intelligence | 27 | 10 |
| AA-LCR | 66 | 20.3 |
| AA-Omniscience | -14.3 | -52.2 |
| aider_polyglot | 82.2 | 7.6 |
| AIR-Bench 2024 | 73.6 | 59 |
| anthropic_red_team | 99.5 | 99.6 |
| Artificial Analysis Coding Index | 46.7 | 11.9 |
| bbq | 96.4 | 95.4 |
| critpt | 2.6 | 0 |
| Fortress | 54.9 | 56.4 |
| GPQA Diamond | 84.4 | 49.1 |
| harmbench | 65.4 | 72.8 |
| HLE | 22.5 | 4.2 |
| IFBench | 49 | 36.9 |
| LiveCodeBench | 82.7 | 55.5 |
| MMLU-Pro | 86 | 71.6 |
| OmniScience Accuracy | 39 | 17.6 |
| OmniScience Non-Hallucination | 12.6 | 15.3 |
| scicode | 43 | 26.7 |
| simple_safety_tests | 97 | 100 |
| simpleqa | 50.8 | 10.3 |
| SWE-bench Verified | 67.2 | 23.8 |
| Terminal-Bench Hard | 26.5 | 4.5 |
| xstest | 98.7 | 97.9 |
| τ²-Bench Telecom (AA run) | 54.1 | 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.