| Benchmark | DeepSeek-V4-Flash | Gemma 4 31B |
|---|---|---|
| AA Agentic Index | 61.3 | 14.4 |
| AA Intelligence | 51.8 | 29.7 |
| AA-LCR | 74.3 | 68.3 |
| AA-Omniscience | -14.3 | -47.9 |
| aime_2026 | 95.8 | 89.2 |
| arena_elo | 1439 | 1451 |
| arena_text_factuality | 1434 | 1449 |
| Artificial Analysis Coding Index | 69.1 | 43.4 |
| coding_arena_elo | 1431 | 1364 |
| critpt | 16.6 | 1.4 |
| gdpval | 52.9 | 15.5 |
| GPQA Diamond | 90.8 | 85.7 |
| HLE | 38.6 | 26.5 |
| HLE (with tools) | 45.1 | 26.5 |
| hmmt_feb_2026 | 94.8 | 77.2 |
| IFBench | 79.2 | 75.6 |
| imo_answer_bench | 88.4 | 74.5 |
| itbenchSre | 31.5 | 37.3 |
| MCP Atlas | 69 | 57.2 |
| MMLU-Pro | 86.4 | 85.2 |
| nl2repo | 54.2 | 15.5 |
| OmniScience Accuracy | 40.4 | 20 |
| OmniScience Non-Hallucination | 15.6 | 18.1 |
| scicode | 49.9 | 43.4 |
| simpleqa_verified | 45 | 9.6 |
| SWE-bench Multilingual | 73.3 | 51.7 |
| SWE-bench Pro | 52.6 | 35.7 |
| SWE-bench Verified | 79 | 52 |
| TauBench V3 - Banking | 39.4 | 14.8 |
| Terminal-Bench 2.0 | 56.9 | 42.9 |
| Terminal-Bench 2.1 | 82.7 | 43.4 |
| Terminal-Bench Hard | 38.6 | 36.4 |
| τ²-Bench Telecom (AA run) | 95.6 | 65.5 |
| τ³-Bench | 22.9 | 67.5 |
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.