Calibration
BrowseComp → GDPval-AA
GDPval-AA is estimated from BrowseComp with a Hill curve fitted on 28 models measured on both: y = 0.1854 + (1.2000 − 0.1854)·x^6.00 / (1.01170^6.00 + x^6.00), R² = 0.70, cross-validated error 8.4 pp. It is used for 7 estimates.
| Estimated model | BrowseComp | GDPval-AA | Source |
|---|---|---|---|
| Agents-A1-4B | 66.8% | 26.3% | estimated ± 8.4 pp, medium confidence |
| GPT-5.4 Pro | 89.3% | 51.1% | estimated ± 8.4 pp, medium confidence |
| GPT-5.5 Pro | 90.1% | 52.3% | estimated ± 8.4 pp, medium confidence |
| LongCat-Flash-Lite-Sparse | 48.6% | 19.8% | estimated ± 8.4 pp, medium confidence |
| Qwen3.5-27B | 61.0% | 23.2% | estimated ± 8.4 pp, medium confidence |
| Qwen3.5-35B-A3B | 61.0% | 23.2% | estimated ± 8.4 pp, medium confidence |
| Beam | 77.4% | 35.5% | estimated ± 8.4 pp, medium confidence |