benchgap
Calibration

CyberGym → AA Agentic Index

AA Agentic Index is estimated from CyberGym with a inverse Michaelis–Menten curve fitted on 10 models measured on both: y = 0.34244·(x − 0.0000) / (0.0000 + 1.4295 − x), R² = 0.78, cross-validated error 5.8 pp. It is used for 6 estimates.

Estimated modelCyberGymAA Agentic IndexSource
Atria Dawn Preview86.5%52.5%estimated ± 5.8 pp, low confidence
Claude Mythos 583.8%48.5%estimated ± 5.8 pp, medium confidence
Claude Mythos Preview83.1%47.5%estimated ± 5.8 pp, medium confidence
Gemini 3.5 Flash Cyber83.2%47.7%estimated ± 5.8 pp, medium confidence
Gemini 3.8 Flash Cyber86.2%52.0%estimated ± 5.8 pp, low confidence
Fugu Cyber86.9%53.1%estimated ± 5.8 pp, low confidence