| Benchmark | Claude Sonnet 4.6 | GPT-4o |
|---|---|---|
| AA Agentic Index | 61.6 | 9.7 |
| AA Intelligence | 48.4 | 11.2 |
| AA-LCR | 74 | 55 |
| AA-Omniscience | 12.2 | -10.5 |
| arc_agi_1 | 86.5 | 4.5 |
| ARC-AGI-2 | 60.4 | 0 |
| arena_vision | 1278 | 1162 |
| Artificial Analysis Coding Index | 63 | 24.2 |
| critpt | 3.1 | 0 |
| gdpval | 54.8 | 0 |
| GPQA Diamond | 89.9 | 70.1 |
| harmbench | 24.2 | 82.9 |
| HLE | 49 | 5.3 |
| IFBench | 56.6 | 36 |
| mmmlu | 89.3 | 81.4 |
| MMMU-Pro | 75.6 | 59.9 |
| OmniScience Accuracy | 40.9 | 23.7 |
| OmniScience Non-Hallucination | 51.6 | 62.1 |
| OSWorld-Verified | 78.5 | 5 |
| scicode | 47 | 33.4 |
| SWE-bench Verified | 79.6 | 38.8 |
| Tau2 retail | 91.7 | 63.4 |
| Terminal-Bench Hard | 59.1 | 8.3 |
| vectara_answer_rate | 99.9 | 93.8 |
| vectara_avg_summary_length | 114.7 | 86.6 |
| vectara_factual_consistency | 89.4 | 90.4 |
| vectara_hallucination_rate ↓ | 10.6 | 9.6 |
| τ²-Bench Telecom (AA run) | 97.9 | 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.