| Benchmark | GPT-5.4 | Qwen3 VL 4B (Reasoning) |
|---|---|---|
| AA Agentic Index | 58.2 | 14.4 |
| AA Intelligence | 53.1 | 7.7 |
| AA-LCR | 77.7 | 23 |
| AA-Omniscience | 5.8 | -68.4 |
| Artificial Analysis Coding Index | 71.1 | 6.7 |
| critpt | 23.4 | 0 |
| gdpval | 50.1 | 13.8 |
| GPQA Diamond | 93 | 64.1 |
| HLE | 43.7 | 4.6 |
| IFBench | 73.9 | 36.6 |
| mathvision | 92 | 60 |
| MMLU-Pro | 87.5 | 73.6 |
| MMMU-Pro | 81.2 | 57 |
| OmniScience Accuracy | 50.9 | 12.2 |
| OmniScience Non-Hallucination | 17.4 | 8.2 |
| OSWorld-Verified | 75 | 31.4 |
| scicode | 56.6 | 17.1 |
| Terminal-Bench Hard | 57.6 | 1.5 |
| τ²-Bench Telecom (AA run) | 87.1 | 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.