| Benchmark | GPT-5 | Qwen2.5 Instruct 72B |
|---|---|---|
| AA Intelligence | 35.3 | 10 |
| AA-LCR | 76.3 | 20.3 |
| AA-Omniscience | -8.7 | -52.2 |
| aider_polyglot | 88 | 7.6 |
| AIR-Bench 2024 | 87.7 | 59 |
| Artificial Analysis Coding Index | 38.9 | 11.9 |
| critpt | 5.7 | 0 |
| Fortress | 17 | 56.4 |
| GPQA Diamond | 87.3 | 49.1 |
| GSM8K | -1.8 | 95.8 |
| harmbench | 70 | 72.8 |
| HLE | 28.5 | 4.2 |
| humaneval | 93.4 | 86.6 |
| IFBench | 73.1 | 36.9 |
| LiveCodeBench | 85 | 55.5 |
| mmlu_redux | 95.3 | 86.8 |
| MMLU-Pro | 87.5 | 71.6 |
| OmniScience Accuracy | 40.3 | 17.6 |
| OmniScience Non-Hallucination | 21 | 15.3 |
| scicode | 43 | 26.7 |
| SWE-bench Verified | 74.9 | 23.8 |
| Terminal-Bench Hard | 37.9 | 4.5 |
| τ²-Bench Telecom (AA run) | 86.5 | 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.