benchgap
Calibration

OSWorld-Verified → ApprenticeBench

ApprenticeBench is estimated from OSWorld-Verified with a linear curve fitted on 6 models measured on both: y = 2.1592·x + -1.5257, R² = 0.90, cross-validated error 5.2 pp. It is used for 11 estimates.

Estimated modelOSWorld-VerifiedApprenticeBenchSource
Gemini 3.5 Flash-Lite74.0%7.2%estimated ± 5.2 pp, low confidence
Gemini 3.6 Flash83.0%26.7%estimated ± 5.2 pp, low confidence
Holo3-122B-A10B78.9%17.7%estimated ± 5.2 pp, low confidence
Holo3-35B-A3B82.6%25.7%estimated ± 5.2 pp, low confidence
MiniMax M370.1%0.0%estimated ± 5.2 pp, low confidence
Muse Glimmer 30B65.9%0.0%estimated ± 5.2 pp, low confidence
Qwen3.5-122B-A10B58.0%0.0%estimated ± 5.2 pp, low confidence
Qwen3.8-27B84.3%29.5%estimated ± 5.2 pp, low confidence
Qwen3.8 Max86.1%33.3%estimated ± 5.2 pp, low confidence
UI-Mate-27B77.0%13.7%estimated ± 5.2 pp, low confidence
UI-Mate-9B66.2%0.0%estimated ± 5.2 pp, low confidence