benchgap
Calibration

OSWorld 2.0 → OSWorld-Verified

OSWorld-Verified is estimated from OSWorld 2.0 with a linear curve fitted on 9 models measured on both: y = 0.7145·x + 0.6889, R² = 0.80, cross-validated error 3.0 pp. It is used for 4 estimates.

Estimated modelOSWorld 2.0OSWorld-VerifiedSource
Claude Opus 5.548.7%100.0%estimated ± 3.0 pp, medium confidence
Gemini 4 Argon69.2%100.0%estimated ± 3.0 pp, medium confidence
GPT-6 Sol60.5%100.0%estimated ± 3.0 pp, medium confidence
Qwen3.8-Flash-Next19.4%82.8%estimated ± 3.0 pp, high confidence