benchgap
Calibration

ExploitBench Internal Port → Chess Puzzles

Chess Puzzles is estimated from ExploitBench Internal Port with a Hill curve fitted on 6 models measured on both: y = 0.3604 + (1.2000 − 0.3604)·x^0.46 / (0.57325^0.46 + x^0.46), R² = 1.00, cross-validated error 5.3 pp. It is used for 3 estimates.

Estimated modelExploitBench Internal PortChess PuzzlesSource
GPT-6.2 Sol50.5%76.8%estimated ± 5.3 pp, low confidence
GPT-6 Cyber42.0%75.0%estimated ± 5.3 pp, low confidence
GPT-6.1 Terra24.5%69.9%estimated ± 5.3 pp, low confidence

Fitted on

ModelExploitBench Internal PortChess Puzzles
GPT-5.6 Sol5.5%59.0%
GPT-6.1 Sol21.5%68.0%
GPT-6.2 Astra53.5%78.0%
GPT-6 Astra39.0%74.0%
GPT-6 Luna0.0%36.0%
GPT-6 Sol5.5%56.0%