| Benchmark | DeepSeek-V3 | DeepSeek-V4-Flash |
|---|---|---|
| AA Agentic Index | 1.6 | 61.3 |
| AA Intelligence | 15.4 | 51.8 |
| AA-LCR | 41.3 | 74.3 |
| AA-Omniscience | -37.6 | -14.3 |
| Artificial Analysis Coding Index | 23 | 69.1 |
| Chinese SimpleQA (C-SimpleQA) | 68 | 78.9 |
| Codeforces | 1134 | 3052 |
| critpt | 0 | 16.6 |
| gdpval | 0 | 52.9 |
| GPQA Diamond | 68.4 | 90.8 |
| HLE | 5.2 | 38.6 |
| IFBench | 41 | 79.2 |
| LiveCodeBench | 49.6 | 91.6 |
| MMLU-Pro | 81.2 | 86.4 |
| OmniScience Accuracy | 25.4 | 40.4 |
| OmniScience Non-Hallucination | 23.3 | 15.6 |
| scicode | 35.8 | 49.9 |
| simplebench | 27.2 | 46.3 |
| SWE-bench Verified | 42 | 79 |
| TauBench V3 - Banking | 4.7 | 39.4 |
| Terminal-Bench 2.1 | 16.9 | 82.7 |
| Terminal-Bench Hard | 15.2 | 38.6 |
| τ²-Bench Telecom (AA run) | 47.1 | 95.6 |
| τ³-Bench | 4.7 | 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.