benchgap
Calibration

ResearchClawBench → OSWorld-Verified

OSWorld-Verified is estimated from ResearchClawBench with a Hill curve fitted on 9 models measured on both: y = 0.0000 + (0.7934 − 0.0000)·x^6.00 / (0.10963^6.00 + x^6.00), R² = 0.55, cross-validated error 5.5 pp. It is used for 2 estimates.

Estimated modelResearchClawBenchOSWorld-VerifiedSource
Grok 4.113.5%61.7%estimated ± 5.5 pp, low confidence
Grok 4.312.4%53.7%estimated ± 5.5 pp, low confidence

Fitted on

ModelResearchClawBenchOSWorld-Verified
Claude Opus 4.619.9%72.7%
Claude Opus 4.821.1%83.4%
Gemini 3.5 Flash18.0%78.4%
GPT-5.415.3%75.0%
GPT-5.517.0%78.7%
Kimi K2.618.0%73.1%
Kimi K2.514.0%63.3%
MiniMax M319.8%70.1%
Qwen3.5 397B14.2%62.2%