| Benchmark | DeepSeek-V3 | GPT-5.2 |
|---|---|---|
| AA Agentic Index | 1.6 | 60.2 |
| AA Intelligence | 15.4 | 43.3 |
| AA-LCR | 41.3 | 79.3 |
| AA-Omniscience | -37.6 | 0.3 |
| AIME 2025 | 51.3 | 100 |
| Artificial Analysis Coding Index | 23 | 48.7 |
| critpt | 0 | 11.6 |
| gdpval | 0 | 48.3 |
| GPQA Diamond | 68.4 | 92.4 |
| HLE | 5.2 | 45.5 |
| HMMT 2025 | 27.5 | 99.4 |
| hmmt_feb_2025 | 29.2 | 100 |
| IFBench | 41 | 75.4 |
| livebench | 72.4 | 74.8 |
| LiveCodeBench | 49.6 | 89 |
| longbench_v2 | 48.7 | 54.5 |
| MMLU-Pro | 81.2 | 87 |
| mmmlu | 79.4 | 89.6 |
| OmniScience Accuracy | 25.4 | 44.3 |
| OmniScience Non-Hallucination | 23.3 | 38.4 |
| scicode | 35.8 | 52.1 |
| simplebench | 27.2 | 45.8 |
| SWE-bench Verified | 42 | 80 |
| Tau2 retail | 69.1 | 82 |
| TauBench V3 - Banking | 4.7 | 11.1 |
| Terminal-Bench Hard | 15.2 | 62.2 |
| vectara_answer_rate | 97.5 | 100 |
| vectara_avg_summary_length | 81.7 | 186.3 |
| vectara_factual_consistency | 93.9 | 89.2 |
| vectara_hallucination_rate ↓ | 6.1 | 10.8 |
| τ²-Bench | 32.5 | 85.5 |
| τ²-Bench Telecom (AA run) | 47.1 | 98.7 |
| τ³-Bench | 4.7 | 12 |
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.