| Benchmark | DeepSeek-V3 | Kimi K2.5 |
|---|---|---|
| AA Agentic Index | 1.6 | 52.8 |
| AA Intelligence | 15.4 | 36 |
| AA-LCR | 41.3 | 73 |
| AA-Omniscience | -37.6 | -7.3 |
| AIME 2025 | 51.3 | 96.1 |
| Artificial Analysis Coding Index | 23 | 46.8 |
| critpt | 0 | 3.1 |
| gdpval | 0 | 38.3 |
| GPQA Diamond | 68.4 | 87.9 |
| HLE | 5.2 | 50.2 |
| HMMT 2025 | 27.5 | 95.4 |
| hmmt_feb_2025 | 29.2 | 95.4 |
| IFBench | 41 | 70.2 |
| livebench | 72.4 | 69.1 |
| LiveCodeBench v6 | 46.9 | 85 |
| longbench_v2 | 48.7 | 61 |
| MMLU-Pro | 81.2 | 87.1 |
| multichallenge | 31.4 | 61.4 |
| OmniScience Accuracy | 25.4 | 35.2 |
| OmniScience Non-Hallucination | 23.3 | 50 |
| scicode | 35.8 | 49 |
| simplebench | 27.2 | 46.8 |
| SWE-bench Verified | 42 | 76.8 |
| TauBench V3 - Banking | 4.7 | 14.2 |
| Terminal-Bench 2.1 | 16.9 | 45.7 |
| Terminal-Bench Hard | 15.2 | 34.8 |
| vectara_answer_rate | 97.5 | 92.2 |
| vectara_avg_summary_length | 81.7 | 112 |
| vectara_factual_consistency | 93.9 | 85.8 |
| vectara_hallucination_rate ↓ | 6.1 | 14.2 |
| τ²-Bench | 32.5 | 85.4 |
| τ²-Bench Telecom (AA run) | 47.1 | 95.9 |
| τ³-Bench | 4.7 | 66 |
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.