| Benchmark | Claude Sonnet 4.6 | DeepSeek-R1 |
|---|---|---|
| AA Agentic Index | 61.6 | 3.1 |
| AA Intelligence | 48.4 | 18.6 |
| AA-LCR | 74 | 56 |
| AA-Omniscience | 12.2 | -31.3 |
| arc_agi_1 | 86.5 | 21.2 |
| ARC-AGI-2 | 60.4 | 1.3 |
| Artificial Analysis Coding Index | 63 | 24.6 |
| browsecomp | 76.2 | 8.9 |
| critpt | 3.1 | 0.6 |
| gdpval | 54.8 | 1.5 |
| GPQA Diamond | 89.9 | 81 |
| harmbench | 24.2 | 54.6 |
| HLE | 49 | 17.7 |
| IFBench | 56.6 | 39 |
| OmniScience Accuracy | 40.9 | 30.7 |
| OmniScience Non-Hallucination | 51.6 | 10.5 |
| scicode | 47 | 35.7 |
| simpleqa_verified | 29 | 27.4 |
| SWE-bench Verified | 79.6 | 57.6 |
| TauBench V3 - Banking | 34.4 | 6.4 |
| Terminal-Bench 2.1 | 71.2 | 19.1 |
| Terminal-Bench Hard | 59.1 | 6.1 |
| vectara_answer_rate | 99.9 | 97 |
| vectara_avg_summary_length | 114.7 | 93.5 |
| vectara_factual_consistency | 89.4 | 88.7 |
| vectara_hallucination_rate ↓ | 10.6 | 11.3 |
| τ²-Bench Telecom (AA run) | 97.9 | 11.4 |
| τ³-Bench | 30.5 | 6.4 |
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.