| Benchmark | GPT-5.2 | Qwen3 VL 4B (Reasoning) |
|---|---|---|
| AA Agentic Index | 60.2 | 14.4 |
| AA Intelligence | 43.3 | 7.7 |
| AA-LCR | 79.3 | 23 |
| AA-Omniscience | 0.3 | -68.4 |
| AIME 2025 | 100 | 74.5 |
| Artificial Analysis Coding Index | 48.7 | 6.7 |
| critpt | 11.6 | 0 |
| gdpval | 48.3 | 13.8 |
| GPQA Diamond | 92.4 | 64.1 |
| HLE | 45.5 | 4.6 |
| HMMT 2025 | 99.4 | 53.1 |
| IFBench | 75.4 | 36.6 |
| InfoVQA (val) | 84 | 79.5 |
| LiveCodeBench | 89 | 51.3 |
| LongVideoBench | 76.5 | 57.7 |
| mathvision | 83 | 60 |
| MathVista | 82.8 | 79.5 |
| MMLU-Pro | 87 | 73.6 |
| MMMU-Pro | 79.5 | 57 |
| OCRBench | 80.7 | 81.6 |
| OmniScience Accuracy | 44.3 | 12.2 |
| OmniScience Non-Hallucination | 38.4 | 8.2 |
| scicode | 52.1 | 17.1 |
| screenspot_pro_no_tools | 86.3 | 49.2 |
| Terminal-Bench Hard | 62.2 | 1.5 |
| Video-MME | 86 | 59.7 |
| VideoMMMU | 85.9 | 69.4 |
| τ²-Bench Telecom (AA run) | 98.7 | 15.5 |
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.