| Benchmark | DeepSeek-R1 | Gemma 4 26B A4B |
|---|---|---|
| AA Agentic Index | 3.1 | 32.1 |
| AA Intelligence | 18.6 | 26.1 |
| AA-LCR | 56 | 61.7 |
| AA-Omniscience | -31.3 | -50.8 |
| AIME 2025 | 87.5 | 90 |
| Artificial Analysis Coding Index | 24.6 | 39.3 |
| browsecomp | 8.9 | 26.3 |
| C-Eval | 91.8 | 82.5 |
| critpt | 0.6 | 0 |
| gdpval | 1.5 | 25.7 |
| GPQA Diamond | 81 | 82.3 |
| HLE | 17.7 | 19.3 |
| hmmt_feb_2025 | 76.7 | 91.7 |
| IFBench | 39 | 72.4 |
| ifeval | 83.3 | 96.4 |
| LiveCodeBench | 84.4 | 79.8 |
| mmlu | 92.9 | 77.8 |
| mmlu_redux | 93.4 | 92.7 |
| MMLU-Pro | 85 | 85.2 |
| OmniScience Accuracy | 30.7 | 19.1 |
| OmniScience Non-Hallucination | 10.5 | 13.6 |
| scicode | 35.7 | 40.3 |
| SWE-bench Multilingual | 30.5 | 43.4 |
| SWE-bench Verified | 57.6 | 57.4 |
| TauBench V3 - Banking | 6.4 | 12 |
| Terminal-Bench 2.1 | 19.1 | 39 |
| Terminal-Bench Hard | 6.1 | 25 |
| τ²-Bench Telecom (AA run) | 11.4 | 43.6 |
| τ³-Bench | 6.4 | 59 |
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.