Calibration
OSWorld 2.0 → AA Harvey LAB
AA Harvey LAB is estimated from OSWorld 2.0 with a linear curve fitted on 5 models measured on both: y = 0.0726·x + 0.8826, R² = 0.73, cross-validated error 1.3 pp. It is used for 6 estimates.
| Estimated model | OSWorld 2.0 | AA Harvey LAB | Source |
|---|---|---|---|
| Claude Opus 4.7 | 13.9% | 89.3% | estimated ± 1.3 pp, medium confidence |
| Claude Sonnet 4.6 | 8.3% | 88.9% | estimated ± 1.3 pp, medium confidence |
| Gemini 4 Argon | 69.2% | 93.3% | estimated ± 1.3 pp, medium confidence |
| GPT-6 Sol | 60.5% | 92.7% | estimated ± 1.3 pp, medium confidence |
| Qwen3.8-Flash-Next | 19.4% | 89.7% | estimated ± 1.3 pp, medium confidence |
| Qwen3.8 Max | 19.4% | 89.7% | estimated ± 1.3 pp, medium confidence |