Calibration

WideResearch → CyberGym

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap
License: https://creativecommons.org/licenses/by-nc/4.0/

CyberGym is estimated from WideResearch with a Hill curve fitted on 5 models measured on both: y = 0.0000 + (1.2000 − 0.0000)·x^6.00 / (0.74868^6.00 + x^6.00), R² = 0.78, cross-validated error 11.1 pp. It is used for 0 estimates.

Fitted on

Model	WideResearch	CyberGym
Atria Dawn Preview	81.9%	86.5%
Claude Opus 4.5	76.4%	50.6%
GLM-5	69.8%	43.2%
Hy4 preview	83.9%	78.4%
Ling 3.1 Flash	83.3%	87.9%