| Benchmark | Gemma 4 31B | GPT-5.5 |
|---|---|---|
| AA Agentic Index | 14.4 | 47.4 |
| AA Intelligence | 29.7 | 56.3 |
| AA-LCR | 68.3 | 79 |
| AA-Omniscience | -47.9 | 20.5 |
| aime_2026 | 89.2 | 100 |
| arena_elo | 1451 | 1482 |
| arena_text_factuality | 1449 | 1482 |
| arena_vision | 1252 | 1283 |
| Artificial Analysis Coding Index | 43.4 | 74.9 |
| coding_arena_elo | 1364 | 1457 |
| critpt | 1.4 | 27.1 |
| gdpval | 15.5 | 49.5 |
| GPQA Diamond | 85.7 | 93.6 |
| HLE | 26.5 | 52.2 |
| HLE (with tools) | 26.5 | 52.2 |
| hmmt_feb_2026 | 77.2 | 98.5 |
| hmmt_nov_2025 | 87.5 | 96.5 |
| IFBench | 75.6 | 75.9 |
| itbenchSre | 37.3 | 45.8 |
| MCP Atlas | 57.2 | 82.8 |
| MMMU-Pro | 76.9 | 83.2 |
| nl2repo | 15.5 | 50.7 |
| OmniScience Accuracy | 20 | 58 |
| OmniScience Non-Hallucination | 18.1 | 12 |
| scicode | 43.4 | 56.1 |
| simpleqa_verified | 9.6 | 63.1 |
| SWE-bench Pro | 35.7 | 58.6 |
| SWE-bench Verified | 52 | 80.6 |
| TauBench V3 - Banking | 14.8 | 39 |
| Terminal-Bench 2.0 | 42.9 | 82.7 |
| Terminal-Bench 2.1 | 43.4 | 84.3 |
| Terminal-Bench Hard | 36.4 | 60.6 |
| tool_decathlon | 21.2 | 55.6 |
| vectara_answer_rate | 100 | 100 |
| vectara_avg_summary_length | 75.8 | 129.6 |
| vectara_factual_consistency | 92.6 | 90.7 |
| vectara_hallucination_rate ↓ | 7.4 | 9.3 |
| τ²-Bench Telecom (AA run) | 65.5 | 93.9 |
| τ³-Bench | 67.5 | 31.3 |
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.