| Benchmark | Claude 4 Opus | Qwen2.5 Instruct 72B |
|---|---|---|
| AA Intelligence | 31.7 | 10 |
| AA-LCR | 40 | 20.3 |
| aider_polyglot | 70.7 | 7.6 |
| AIME 2024 | 76 | 23.3 |
| AIR-Bench 2024 | 85.7 | 59 |
| anthropic_red_team | 98.9 | 99.6 |
| Artificial Analysis Coding Index | 34 | 11.9 |
| bbq | 99.3 | 95.4 |
| Fortress | 27.6 | 56.4 |
| GPQA Diamond | 79.6 | 49.1 |
| harmbench | 91.7 | 72.8 |
| HLE | 12.3 | 4.2 |
| IFBench | 53.7 | 36.9 |
| ifeval | 87.4 | 87.2 |
| livebench | 74.6 | 52.3 |
| LiveCodeBench | 70.4 | 55.5 |
| longbench_v2 | 55.6 | 39.4 |
| math_500_em | 98.2 | 80 |
| mmlu | 92.9 | 86.1 |
| mmlu_redux | 94.2 | 86.8 |
| MMLU-Pro | 86.6 | 71.6 |
| mmmlu | 88.8 | 74.8 |
| MultiPL-E | 89.6 | 75.1 |
| scicode | 40.9 | 26.7 |
| simple_safety_tests | 100 | 100 |
| simpleqa | 22.8 | 10.3 |
| SWE-bench Verified | 79.4 | 23.8 |
| Terminal-Bench Hard | 31.1 | 4.5 |
| xstest | 97 | 97.9 |
| τ²-Bench Telecom (AA run) | 73.4 | 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.