| Benchmark | Claude Opus 4.6 | Qwen3 VL 4B (Reasoning) |
|---|---|---|
| AA Agentic Index | 67.6 | 14.4 |
| AA Intelligence | 44.9 | 7.7 |
| AA-LCR | 74.3 | 23 |
| AA-Omniscience | 13.7 | -68.4 |
| AIME 2025 | 99.8 | 74.5 |
| Artificial Analysis Coding Index | 48.1 | 6.7 |
| critpt | 12.6 | 0 |
| ERQA | 40.8 | 47.3 |
| gdpval | 55.9 | 13.8 |
| GPQA Diamond | 91.3 | 64.1 |
| HLE | 53.1 | 4.6 |
| IFBench | 62.5 | 36.6 |
| LiveCodeBench | 88.8 | 51.3 |
| LiveCodeBench v6 | 88.8 | 51.3 |
| mathvision | 71.2 | 60 |
| MMLU-Pro | 89.1 | 73.6 |
| MMMU-Pro | 77.3 | 57 |
| OmniScience Accuracy | 47 | 12.2 |
| OmniScience Non-Hallucination | 37.2 | 8.2 |
| OSWorld-Verified | 72.7 | 31.4 |
| RealWorldQA | 73.9 | 73.2 |
| scicode | 52 | 17.1 |
| screenspot_pro_no_tools | 57.7 | 49.2 |
| Terminal-Bench Hard | 65.4 | 1.5 |
| τ²-Bench Telecom (AA run) | 99.3 | 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.