| Benchmark | Claude Sonnet 4 | Qwen3 VL 4B (Reasoning) |
|---|---|---|
| AA Agentic Index | 39.2 | 14.4 |
| AA Intelligence | 29.8 | 7.7 |
| AA-LCR | 69.7 | 23 |
| AA-Omniscience | 0.2 | -68.4 |
| AIME 2024 | 43.4 | 72.9 |
| AIME 2025 | 74 | 74.5 |
| Artificial Analysis Coding Index | 37.6 | 6.7 |
| critpt | 1.1 | 0 |
| gdpval | 31.2 | 13.8 |
| GPQA Diamond | 78 | 64.1 |
| HLE | 10.7 | 4.6 |
| HMMT 2025 | 15.9 | 53.1 |
| IFBench | 55 | 36.6 |
| ifeval | 87.6 | 82.6 |
| LiveCodeBench | 68.5 | 51.3 |
| LiveCodeBench v6 | 48.5 | 51.3 |
| mmlu_redux | 93.6 | 86 |
| MMLU-Pro | 84 | 73.6 |
| MMMU-Pro | 62.4 | 57 |
| OmniScience Accuracy | 22.7 | 12.2 |
| OmniScience Non-Hallucination | 70.9 | 8.2 |
| OSWorld-Verified | 42.2 | 31.4 |
| scicode | 40 | 17.1 |
| supergpqa | 55.7 | 46.8 |
| Terminal-Bench Hard | 31.1 | 1.5 |
| τ²-Bench Telecom (AA run) | 64.6 | 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.