| Benchmark | Gemma 4 31B | GPT-5.6 Sol |
|---|---|---|
| AA Agentic Index | 14.4 | 57.8 |
| AA Intelligence | 29.7 | 61 |
| AA-LCR | 68.3 | 77.7 |
| AA-Omniscience | -47.9 | 22 |
| aime_2026 | 89.2 | 99.9 |
| arena_elo | 1451 | 1482 |
| arena_text_factuality | 1449 | 1474 |
| Artificial Analysis Coding Index | 43.4 | 78.3 |
| charxiv_rq | 67.9 | 87.8 |
| coding_arena_elo | 1364 | 1619 |
| critpt | 1.4 | 32.3 |
| gdpval | 15.5 | 61.1 |
| GPQA Diamond | 85.7 | 94.6 |
| HLE | 26.5 | 49.5 |
| HLE (with tools) | 26.5 | 58 |
| IFBench | 75.6 | 72.7 |
| itbenchSre | 37.3 | 56.2 |
| mathvision | 85.6 | 95.8 |
| MCP Atlas | 57.2 | 83.6 |
| MMMU-Pro | 76.9 | 84.6 |
| OmniScience Accuracy | 20 | 59.4 |
| OmniScience Non-Hallucination | 18.1 | 10.6 |
| scicode | 43.4 | 56.9 |
| simpleqa_verified | 9.6 | 71.6 |
| SWE-bench Pro | 35.7 | 64.6 |
| TauBench V3 - Banking | 14.8 | 44.3 |
| Terminal-Bench 2.1 | 43.4 | 89.5 |
| Terminal-Bench Hard | 36.4 | 65.9 |
| τ²-Bench Telecom (AA run) | 65.5 | 85.1 |
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.