| Benchmark | DeepSeek-V3.2 | Gemini 2.5 Flash |
|---|---|---|
| AA Agentic Index | 39.8 | 18.8 |
| AA Intelligence | 32.8 | 17.7 |
| AA-LCR | 70.7 | 65.7 |
| AA-Omniscience | -22.5 | -29.8 |
| AIME 2025 | 93.1 | 72 |
| arc_agi_1 | 57 | 33.3 |
| ARC-AGI-2 | 4 | 2.5 |
| arena_text_factuality | 1433 | 1437 |
| Artificial Analysis Coding Index | 44.2 | 22.2 |
| critpt | 2.9 | 1.4 |
| frontiermath_tier_4 | 2.1 | 4.2 |
| gdpval | 18.8 | 11.9 |
| GPQA Diamond | 84 | 82.8 |
| HLE | 40.8 | 12.1 |
| hmmt_feb_2025 | 92.5 | 64.2 |
| IFBench | 61 | 50.3 |
| LiveCodeBench | 86 | 76.2 |
| mrcr | 55.5 | 32 |
| OmniScience Accuracy | 33 | 26.1 |
| OmniScience Non-Hallucination | 17.3 | 24.7 |
| scicode | 39 | 35.9 |
| swe_bench_bash | 60 | 28.7 |
| SWE-bench Verified | 73.1 | 60.4 |
| Terminal-Bench Hard | 35.6 | 13.6 |
| vectara_answer_rate | 92.6 | 99 |
| vectara_avg_summary_length | 62 | 101.5 |
| vectara_factual_consistency | 93.7 | 92.2 |
| vectara_hallucination_rate ↓ | 6.3 | 7.8 |
| τ²-Bench Telecom (AA run) | 90.6 | 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.