| Benchmark | GPT-4o mini | Qwen2.5 Instruct 72B |
|---|---|---|
| AA Intelligence | 7 | 10 |
| aider_polyglot | 3.6 | 7.6 |
| AIR-Bench 2024 | 56.3 | 59 |
| anthropic_red_team | 98.3 | 99.6 |
| Artificial Analysis Coding Index | 11.4 | 11.9 |
| bbq | 88.2 | 95.4 |
| DROP | 79.7 | 76.7 |
| Fortress | 48.1 | 56.4 |
| GPQA Diamond | 42.6 | 49.1 |
| GSM8K | 84.3 | 95.8 |
| harmbench | 84.9 | 72.8 |
| HLE | 4.2 | 4.2 |
| humaneval | 87.2 | 86.6 |
| IFBench | 31 | 36.9 |
| LiveCodeBench | 35.5 | 55.5 |
| MATH | 80.2 | 88.4 |
| mgsm | 87 | 87.3 |
| mmlu | 81.8 | 86.1 |
| narrativeqa | 76.8 | 74.5 |
| naturalquestions_closedbook | 38.5 | 35.9 |
| OpenBookQA | 92 | 96.2 |
| scicode | 22.9 | 26.7 |
| simple_safety_tests | 97.8 | 100 |
| simpleqa | 9.9 | 10.3 |
| SWE-bench Verified | 8.7 | 23.8 |
| xstest | 96 | 97.9 |
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.