| Benchmark | DeepSeek-V3 | Gemini 2.5 Flash |
|---|---|---|
| AA Agentic Index | 1.6 | 18.8 |
| AA Intelligence | 15.4 | 17.7 |
| AA-LCR | 41.3 | 65.7 |
| AA-Omniscience | -37.6 | -29.8 |
| aider_polyglot | 55.1 | 61.9 |
| AIME 2024 | 59.4 | 82.3 |
| AIME 2025 | 51.3 | 72 |
| AIR-Bench 2024 | 40.8 | 68.7 |
| anthropic_red_team | 97.1 | 98.8 |
| Artificial Analysis Coding Index | 23 | 22.2 |
| bbq | 96.7 | 97.7 |
| critpt | 0 | 1.4 |
| gdpval | 0 | 11.9 |
| GPQA Diamond | 68.4 | 82.8 |
| harmbench | 49.7 | 62.6 |
| HLE | 5.2 | 12.1 |
| hmmt_feb_2025 | 29.2 | 64.2 |
| IFBench | 41 | 50.3 |
| LiveCodeBench | 49.6 | 76.2 |
| livecodebench_easy | 83.3 | 99.1 |
| livecodebench_hard | 14.2 | 48.3 |
| livecodebench_medium | 53.5 | 82.5 |
| OmniScience Accuracy | 25.4 | 26.1 |
| OmniScience Non-Hallucination | 23.3 | 24.7 |
| scicode | 35.8 | 35.9 |
| simple_safety_tests | 95.3 | 98 |
| simplebench | 27.2 | 41.2 |
| simpleqa | 27.7 | 26.9 |
| SWE-bench Verified | 42 | 60.4 |
| Terminal-Bench Hard | 15.2 | 13.6 |
| vectara_answer_rate | 97.5 | 99 |
| vectara_avg_summary_length | 81.7 | 101.5 |
| vectara_factual_consistency | 93.9 | 92.2 |
| vectara_hallucination_rate ↓ | 6.1 | 7.8 |
| xstest | 97.1 | 98.8 |
| τ²-Bench Telecom (AA run) | 47.1 | 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.