benchgap
Calibration

CyberGym → MCP Atlas

MCP Atlas is estimated from CyberGym with a Hill curve fitted on 12 models measured on both: y = 0.0000 + (0.8126 − 0.0000)·x^6.00 / (0.46036^6.00 + x^6.00), R² = 0.75, cross-validated error 11.1 pp. It is used for 4 estimates.

Estimated modelCyberGymMCP AtlasSource
Fugu Cyber86.9%79.5%estimated ± 11.1 pp, low confidence
Gemini 3.8 Flash Cyber86.2%79.4%estimated ± 11.1 pp, low confidence
Gemini 3.5 Flash Cyber83.2%79.0%estimated ± 11.1 pp, low confidence
Claude Mythos Preview83.1%79.0%estimated ± 11.1 pp, low confidence

Fitted on

ModelCyberGymMCP Atlas
Claude Opus 4.550.6%42.3%
Claude Opus 4.773.1%77.3%
DeepSeek V4 Flash 073176.7%69.0%
DeepSeek V4 Pro83.3%73.6%
GLM-543.2%31.1%
GLM-5.168.7%71.8%
GPT-5.479.0%70.6%
GPT-5.581.8%75.3%
Hy4 preview78.4%83.7%
Ling 3.1 Flash87.9%83.0%
Muse Spark 1.159.0%88.1%
Step 5 Preview84.7%85.6%