| Benchmark | Claude Sonnet 4.5 | Qwen3 VL 4B (Reasoning) |
|---|---|---|
| AA Agentic Index | 50.6 | 14.4 |
| AA Intelligence | 37.4 | 7.7 |
| AA-LCR | 68.3 | 23 |
| AA-Omniscience | -0.1 | -68.4 |
| AI2D | 87 | 84.9 |
| AIME 2025 | 88 | 74.5 |
| Artificial Analysis Coding Index | 52.1 | 6.7 |
| critpt | 1.1 | 0 |
| gdpval | 40.4 | 13.8 |
| GPQA Diamond | 83.4 | 64.1 |
| HallusionBench | 59.9 | 64.1 |
| HLE | 19.8 | 4.6 |
| HMMT 2025 | 74.6 | 53.1 |
| IFBench | 57.3 | 36.6 |
| LiveCodeBench | 71 | 51.3 |
| LiveCodeBench v6 | 64 | 51.3 |
| MathVista | 79.8 | 79.5 |
| mmlu_redux | 95.6 | 86 |
| MMLU-Pro | 88.2 | 73.6 |
| MMMU (val) (Pass@1) | 77.8 | 70.8 |
| MMMU-Pro | 68.7 | 57 |
| OmniScience Accuracy | 32.9 | 12.2 |
| OmniScience Non-Hallucination | 50.9 | 8.2 |
| OSWorld-Verified | 61.4 | 31.4 |
| RealWorldQA | 70.3 | 73.2 |
| scicode | 45 | 17.1 |
| Terminal-Bench Hard | 35.6 | 1.5 |
| τ²-Bench Telecom (AA run) | 78.1 | 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.