| Benchmark | DeepSeek-V3.2 | o4-mini |
|---|---|---|
| AA Agentic Index | 39.8 | 36.1 |
| AA Intelligence | 32.8 | 26.1 |
| AA-LCR | 70.7 | 60 |
| AA-Omniscience | -22.5 | -35.7 |
| AIME 2025 | 93.1 | 92.7 |
| arc_agi_1 | 57 | 58.7 |
| ARC-AGI-2 | 4 | 6.1 |
| Artificial Analysis Coding Index | 44.2 | 25.6 |
| browsecomp | 51.4 | 51.5 |
| critpt | 2.9 | 0.6 |
| frontiermath_tier_4 | 2.1 | 6.3 |
| gdpval | 18.8 | 25.4 |
| GPQA Diamond | 84 | 81.4 |
| HLE | 40.8 | 16.5 |
| IFBench | 61 | 68.7 |
| LiveCodeBench | 86 | 84.5 |
| OmniScience Accuracy | 33 | 24.8 |
| OmniScience Non-Hallucination | 17.3 | 19.5 |
| scicode | 39 | 46.5 |
| simpleqa_verified | 27.5 | 23.9 |
| swe_bench_bash | 60 | 45 |
| SWE-bench Verified | 73.1 | 68.1 |
| Terminal-Bench Hard | 35.6 | 15.2 |
| vectara_answer_rate | 92.6 | 99.2 |
| vectara_avg_summary_length | 62 | 130.9 |
| vectara_factual_consistency | 93.7 | 81.4 |
| vectara_hallucination_rate ↓ | 6.3 | 18.6 |
| τ²-Bench Telecom (AA run) | 90.6 | 55.6 |
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.