benchgap
Calibration

EBR-bench → ExploitBench

ExploitBench is estimated from EBR-bench with a inverse Michaelis–Menten curve fitted on 5 models measured on both: y = 3.32881·(x − 0.3317) / (0.3317 + 0.6428 − x), R² = 0.98, cross-validated error 6.4 pp. It is used for 17 estimates.

Estimated modelEBR-benchExploitBenchSource
Claude Fable 5.157.1%100.0%estimated ± 6.4 pp, low confidence
Claude Fable 5.290.5%100.0%estimated ± 6.4 pp, low confidence
Claude Haiku 5.543.8%65.9%estimated ± 6.4 pp, low confidence
Claude Fable 539.5%36.3%estimated ± 6.4 pp, low confidence
Claude Opus 4.514.3%0.0%estimated ± 6.4 pp, low confidence
Claude Opus 4.612.7%0.0%estimated ± 6.4 pp, low confidence
Claude Opus 4.719.0%0.0%estimated ± 6.4 pp, low confidence
Claude Sonnet 4.52.4%0.0%estimated ± 6.4 pp, low confidence
Gemini 3.5 Flash4.8%0.0%estimated ± 6.4 pp, low confidence
GLM-5.29.5%0.0%estimated ± 6.4 pp, low confidence
GPT-512.7%0.0%estimated ± 6.4 pp, low confidence
GPT-5.223.0%0.0%estimated ± 6.4 pp, low confidence
GPT-5.425.4%0.0%estimated ± 6.4 pp, low confidence
Grok 4.630.5%0.0%estimated ± 6.4 pp, low confidence
Kimi K2.62.4%0.0%estimated ± 6.4 pp, low confidence
Muse Wonder31.4%0.0%estimated ± 6.4 pp, low confidence
Qwen3.7 Max9.5%0.0%estimated ± 6.4 pp, low confidence

Fitted on

ModelEBR-benchExploitBench
GPT-5.6 Sol44.8%73.5%
GPT-6.1 Sol54.3%99.7%
GPT-6.1 Terra78.1%97.1%
GPT-6.2 Astra92.4%100.0%
GPT-6 Astra76.2%100.0%