| Benchmark | Claude Sonnet 4.6 | Gemini 2.5 Flash |
|---|---|---|
| AA Agentic Index | 61.6 | 18.8 |
| AA Intelligence | 48.4 | 17.7 |
| AA-LCR | 74 | 65.7 |
| AA-Omniscience | 12.2 | -29.8 |
| arc_agi_1 | 86.5 | 33.3 |
| ARC-AGI-2 | 60.4 | 2.5 |
| arena_text_factuality | 1460 | 1437 |
| arena_vision | 1278 | 1214 |
| Artificial Analysis Coding Index | 63 | 22.2 |
| critpt | 3.1 | 1.4 |
| frontiermath_tier_4 | 8.3 | 4.2 |
| gdpval | 54.8 | 11.9 |
| GPQA Diamond | 89.9 | 82.8 |
| harmbench | 24.2 | 62.6 |
| HLE | 49 | 12.1 |
| IFBench | 56.6 | 50.3 |
| MMMU-Pro | 75.6 | 69.1 |
| OmniScience Accuracy | 40.9 | 26.1 |
| OmniScience Non-Hallucination | 51.6 | 24.7 |
| scicode | 47 | 35.9 |
| SWE-bench Verified | 79.6 | 60.4 |
| Terminal-Bench Hard | 59.1 | 13.6 |
| vectara_answer_rate | 99.9 | 99 |
| vectara_avg_summary_length | 114.7 | 101.5 |
| vectara_factual_consistency | 89.4 | 92.2 |
| vectara_hallucination_rate ↓ | 10.6 | 7.8 |
| τ²-Bench Telecom (AA run) | 97.9 | 31.6 |
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.