| Benchmark | Claude Sonnet 4 | GPT-4o mini |
|---|---|---|
| AA Agentic Index | 39.2 | 1 |
| AA Intelligence | 29.8 | 7 |
| aider_polyglot | 70.7 | 3.6 |
| AIR-Bench 2024 | 88.3 | 56.3 |
| anthropic_red_team | 99.2 | 98.3 |
| ARC-AGI-2 | 13.6 | 0 |
| Artificial Analysis Coding Index | 37.6 | 11.4 |
| bbq | 97.9 | 88.2 |
| Fortress | 24.4 | 48.1 |
| gdpval | 31.2 | 0 |
| GPQA Diamond | 78 | 42.6 |
| harmbench | 98 | 84.9 |
| HLE | 10.7 | 4.2 |
| IFBench | 55 | 31 |
| LiveCodeBench | 68.5 | 35.5 |
| livecodebench_easy | 98.8 | 81.7 |
| livecodebench_hard | 32.6 | 4.3 |
| livecodebench_medium | 76 | 25.2 |
| mmlu | 91.5 | 81.8 |
| MMMU | 74.4 | 59.4 |
| MMMU-Pro | 62.4 | 41.5 |
| scicode | 40 | 22.9 |
| simple_safety_tests | 100 | 97.8 |
| simplebench | 45.5 | 10.7 |
| simpleqa | 15.9 | 9.9 |
| SWE-bench Verified | 80.2 | 8.7 |
| TauBench V3 - Banking | 16.7 | 2.9 |
| Terminal-Bench 2.1 | 36.3 | 5.6 |
| xstest | 97.2 | 96 |
| τ³-Bench | 13.8 | 2.9 |
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.