| Benchmark | DeepSeek-V4-Pro | Gemini 2.5 Flash |
|---|---|---|
| AA Agentic Index | 63.3 | 18.8 |
| AA Intelligence | 53 | 17.7 |
| AA-LCR | 70 | 65.7 |
| AA-Omniscience | -10.6 | -29.8 |
| arena_text_factuality | 1450 | 1437 |
| Artificial Analysis Coding Index | 59.4 | 22.2 |
| critpt | 13 | 1.4 |
| gdpval | 49 | 11.9 |
| Global-MMLU-Lite | 89.3 | 88.4 |
| GPQA Diamond | 90.5 | 82.8 |
| HLE | 48.2 | 12.1 |
| IFBench | 76.5 | 50.3 |
| LiveCodeBench | 93.5 | 76.2 |
| OmniScience Accuracy | 43 | 26.1 |
| OmniScience Non-Hallucination | 12.2 | 24.7 |
| scicode | 50 | 35.9 |
| simplebench | 50.9 | 41.2 |
| simpleqa | 57.9 | 26.9 |
| SWE-bench Verified | 80.6 | 60.4 |
| Terminal-Bench Hard | 46.2 | 13.6 |
| vectara_answer_rate | 97.2 | 99 |
| vectara_avg_summary_length | 153.8 | 101.5 |
| vectara_factual_consistency | 91.4 | 92.2 |
| vectara_hallucination_rate ↓ | 8.6 | 7.8 |
| τ²-Bench Telecom (AA run) | 96.2 | 31.6 |
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.