Calibration
OSWorld 2.0 → AA AutomationBench
AA AutomationBench is estimated from OSWorld 2.0 with a Michaelis–Menten curve fitted on 8 models measured on both: y = 0.7866·x / (0.12340 + x), R² = 0.86, cross-validated error 7.0 pp. It is used for 1 estimate.
| Estimated model | OSWorld 2.0 | AA AutomationBench | Source |
|---|---|---|---|
| Claude Opus 4.7 | 13.9% | 41.7% | estimated ± 7.0 pp, medium confidence |