benchgap
Calibration

ResearchClawBench → OSWorld 2.0

OSWorld 2.0 is estimated from ResearchClawBench with a offset logistic curve fitted on 5 models measured on both: y = 0.0719 + (1.2000 − 0.0719) / (1 + exp(−200.00·(x − 0.2211))), R² = 0.71, cross-validated error 5.1 pp. It is used for 3 estimates.

Estimated modelResearchClawBenchOSWorld 2.0Source
GLM-5.220.7%13.6%estimated ± 5.1 pp, low confidence
MiMo-V2-Pro15.3%7.2%estimated ± 5.1 pp, low confidence
Grok 4.113.5%7.2%estimated ± 5.1 pp, low confidence

Fitted on

ModelResearchClawBenchOSWorld 2.0
Claude Opus 4.720.7%13.9%
Claude Opus 4.821.1%20.6%
GPT-5.517.0%13.0%
Kimi K2.618.0%4.6%
MiniMax M319.8%4.6%