| Benchmark | Claude Sonnet 4 | Llama 3.1 Instruct 405B |
|---|---|---|
| AA Agentic Index | 39.2 | 6.3 |
| AA Intelligence | 29.8 | 8.5 |
| AA-LCR | 69.7 | 24.7 |
| AA-Omniscience | 0.2 | -17.1 |
| aider_polyglot | 70.7 | 5.8 |
| AIME 2024 | 43.4 | 23.3 |
| AIR-Bench 2024 | 88.3 | 58.6 |
| anthropic_red_team | 99.2 | 96.5 |
| Artificial Analysis Coding Index | 37.6 | 14.5 |
| bbq | 97.9 | 94.5 |
| critpt | 1.1 | 0 |
| Fortress | 24.4 | 20.6 |
| gdpval | 31.2 | 0 |
| GPQA Diamond | 78 | 51.5 |
| harmbench | 98 | 62.7 |
| HLE | 10.7 | 4.2 |
| IFBench | 55 | 39 |
| ifeval | 87.6 | 88.6 |
| LiveCodeBench | 68.5 | 30.1 |
| math_500_em | 94 | 73.8 |
| mmlu | 91.5 | 88.6 |
| mmlu_redux | 93.6 | 86.2 |
| MMLU-Pro | 84 | 73.4 |
| mmmlu | 86.5 | 73.8 |
| OmniScience Accuracy | 22.7 | 23.2 |
| OmniScience Non-Hallucination | 70.9 | 49 |
| scicode | 40 | 29.9 |
| simple_safety_tests | 100 | 98.8 |
| simplebench | 45.5 | 23 |
| simpleqa | 15.9 | 23.2 |
| SWE-bench Pro | 42.7 | 11.2 |
| SWE-bench Verified | 80.2 | 24.5 |
| Terminal-Bench Hard | 31.1 | 6.8 |
| xstest | 97.2 | 95.9 |
| τ²-Bench Telecom (AA run) | 64.6 | 19 |
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.