| Benchmark | DeepSeek-V4-Flash | Qwen3.5 27B |
|---|---|---|
| AA Agentic Index | 61.3 | 54.6 |
| AA Intelligence | 51.8 | 34.6 |
| AA-LCR | 74.3 | 72.3 |
| AA-Omniscience | -14.3 | -44 |
| aime_2026 | 95.8 | 92.6 |
| Artificial Analysis Coding Index | 69.1 | 34.9 |
| browsecomp | 73.2 | 61 |
| coding_arena_elo | 1431 | 1357 |
| critpt | 16.6 | 0.9 |
| gdpval | 52.9 | 33 |
| GPQA Diamond | 90.8 | 85.8 |
| HLE | 38.6 | 48.5 |
| hmmt_feb_2026 | 94.8 | 84.3 |
| IFBench | 79.2 | 76.5 |
| imo_answer_bench | 88.4 | 79.9 |
| itbenchSre | 31.5 | 35.5 |
| MCP Atlas | 69 | 68.4 |
| MMLU-Pro | 86.4 | 86.1 |
| nl2repo | 54.2 | 27.3 |
| OmniScience Accuracy | 40.4 | 20.7 |
| OmniScience Non-Hallucination | 15.6 | 24.9 |
| scicode | 49.9 | 39.5 |
| SWE-bench Multilingual | 73.3 | 69.3 |
| SWE-bench Pro | 52.6 | 51.2 |
| SWE-bench Verified | 79 | 75 |
| Terminal-Bench 2.0 | 56.9 | 41.6 |
| Terminal-Bench Hard | 38.6 | 32.6 |
| τ²-Bench Telecom (AA run) | 95.6 | 93.9 |
| τ³-Bench | 22.9 | 68.4 |
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.