| Benchmark | Claude 4 Opus | DeepSeek-V3.2 |
|---|---|---|
| AA Intelligence | 31.7 | 32.8 |
| AA-LCR | 40 | 70.7 |
| AIME 2025 | 75.5 | 93.1 |
| arc_agi_1 | 27 | 57 |
| ARC-AGI-2 | 8.6 | 4 |
| Artificial Analysis Coding Index | 34 | 44.2 |
| frontiermath_tier_4 | 4.2 | 2.1 |
| GPQA Diamond | 79.6 | 84 |
| HLE | 12.3 | 40.8 |
| HMMT 2025 | 15.9 | 90.2 |
| hmmt_feb_2025 | 60 | 92.5 |
| IFBench | 53.7 | 61 |
| LiveCodeBench | 70.4 | 86 |
| LiveCodeBench v6 | 47.4 | 83.3 |
| longbench_v2 | 55.6 | 59.8 |
| mmlu_redux | 94.2 | 93.7 |
| MMLU-Pro | 86.6 | 86 |
| scicode | 40.9 | 39 |
| swe_bench_bash | 67.6 | 60 |
| SWE-bench Verified | 79.4 | 73.1 |
| Terminal-Bench | 43.2 | 37.7 |
| Terminal-Bench Hard | 31.1 | 35.6 |
| vectara_answer_rate | 91 | 92.6 |
| vectara_avg_summary_length | 123.2 | 62 |
| vectara_factual_consistency | 88 | 93.7 |
| vectara_hallucination_rate ↓ | 12 | 6.3 |
| τ²-Bench | 57 | 91 |
| τ²-Bench Telecom (AA run) | 73.4 | 90.6 |
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.