| Benchmark | Gemma 4 31B | GPT-4o |
|---|---|---|
| AA Agentic Index | 14.4 | 9.7 |
| AA Intelligence | 29.7 | 11.2 |
| AA-LCR | 68.3 | 55 |
| AA-Omniscience | -47.9 | -10.5 |
| AI2D | 89 | 94.2 |
| arena_vision | 1252 | 1162 |
| Artificial Analysis Coding Index | 43.4 | 24.2 |
| C-Eval | 82.6 | 76 |
| critpt | 1.4 | 0 |
| ERQA | 57.5 | 35.2 |
| gdpval | 15.5 | 0 |
| GPQA Diamond | 85.7 | 70.1 |
| HLE | 26.5 | 5.3 |
| IFBench | 75.6 | 36 |
| LiveCodeBench v6 | 80 | 30.9 |
| mathvision | 85.6 | 30.4 |
| MathVista | 79.3 | 63.8 |
| mmlu_redux | 93.7 | 88 |
| MMLU-Pro | 85.2 | 74.7 |
| mmmlu | 88.4 | 81.4 |
| MMMU | 80.4 | 72.2 |
| MMMU-Pro | 76.9 | 59.9 |
| MMStar | 77.3 | 63.9 |
| OCRBench | 86.1 | 806 |
| OmniScience Accuracy | 20 | 23.7 |
| OmniScience Non-Hallucination | 18.1 | 62.1 |
| RealWorldQA | 72.3 | 75.4 |
| scicode | 43.4 | 33.4 |
| supergpqa | 65.7 | 42.4 |
| SWE-bench Verified | 52 | 38.8 |
| Terminal-Bench Hard | 36.4 | 8.3 |
| vectara_answer_rate | 100 | 93.8 |
| vectara_avg_summary_length | 75.8 | 86.6 |
| vectara_factual_consistency | 92.6 | 90.4 |
| vectara_hallucination_rate ↓ | 7.4 | 9.6 |
| τ²-Bench Telecom (AA run) | 65.5 | 28.9 |
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.