| Benchmark | Claude Sonnet 4 | DeepSeek-V4-Flash |
|---|---|---|
| AA Agentic Index | 39.2 | 61.3 |
| AA Intelligence | 29.8 | 51.8 |
| AA-LCR | 69.7 | 74.3 |
| AA-Omniscience | 0.2 | -14.3 |
| arc_agi_1 | 63.7 | 84 |
| ARC-AGI-2 | 13.6 | 46 |
| Artificial Analysis Coding Index | 37.6 | 69.1 |
| browsecomp | 12.2 | 73.2 |
| critpt | 1.1 | 16.6 |
| gdpval | 31.2 | 52.9 |
| GPQA Diamond | 78 | 90.8 |
| HLE | 10.7 | 38.6 |
| HLE (with tools) | 20.3 | 45.1 |
| IFBench | 55 | 79.2 |
| LiveCodeBench | 68.5 | 91.6 |
| MMLU-Pro | 84 | 86.4 |
| OmniScience Accuracy | 22.7 | 40.4 |
| OmniScience Non-Hallucination | 70.9 | 15.6 |
| scicode | 40 | 49.9 |
| simplebench | 45.5 | 46.3 |
| SWE-bench Multilingual | 56.9 | 73.3 |
| SWE-bench Pro | 42.7 | 52.6 |
| SWE-bench Verified | 80.2 | 79 |
| TauBench V3 - Banking | 16.7 | 39.4 |
| Terminal-Bench 2.1 | 36.3 | 82.7 |
| Terminal-Bench Hard | 31.1 | 38.6 |
| τ²-Bench Telecom (AA run) | 64.6 | 95.6 |
| τ³-Bench | 13.8 | 22.9 |
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.