| Benchmark | Claude Sonnet 4 | Gemma 4 31B |
|---|---|---|
| AA Agentic Index | 39.2 | 14.4 |
| AA Intelligence | 29.8 | 29.7 |
| AA-LCR | 69.7 | 68.3 |
| AA-Omniscience | 0.2 | -47.9 |
| arena_vision | 1207 | 1252 |
| Artificial Analysis Coding Index | 37.6 | 43.4 |
| critpt | 1.1 | 1.4 |
| gdpval | 31.2 | 15.5 |
| GPQA Diamond | 78 | 85.7 |
| HLE | 10.7 | 26.5 |
| HLE (with tools) | 20.3 | 26.5 |
| IFBench | 55 | 75.6 |
| LiveCodeBench v6 | 48.5 | 80 |
| mmlu_redux | 93.6 | 93.7 |
| MMLU-Pro | 84 | 85.2 |
| mmmlu | 86.5 | 88.4 |
| MMMU | 74.4 | 80.4 |
| MMMU-Pro | 62.4 | 76.9 |
| OmniScience Accuracy | 22.7 | 20 |
| OmniScience Non-Hallucination | 70.9 | 18.1 |
| scicode | 40 | 43.4 |
| supergpqa | 55.7 | 65.7 |
| SWE-bench Multilingual | 56.9 | 51.7 |
| SWE-bench Pro | 42.7 | 35.7 |
| SWE-bench Verified | 80.2 | 52 |
| TauBench V3 - Banking | 16.7 | 14.8 |
| Terminal-Bench 2.1 | 36.3 | 43.4 |
| Terminal-Bench Hard | 31.1 | 36.4 |
| vectara_answer_rate | 98.6 | 100 |
| vectara_avg_summary_length | 145.8 | 75.8 |
| vectara_factual_consistency | 89.7 | 92.6 |
| vectara_hallucination_rate ↓ | 10.3 | 7.4 |
| τ²-Bench | 65 | 76.9 |
| τ²-Bench Telecom (AA run) | 64.6 | 65.5 |
| τ³-Bench | 13.8 | 67.5 |
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.