benchgap
Calibration

HLE w/ tools → CyberGym

CyberGym is estimated from HLE w/ tools with a offset logistic curve fitted on 5 models measured on both: y = 0.7624 + (1.1656 − 0.7624) / (1 + exp(−20.43·(x − 0.6842))), R² = 0.97, cross-validated error 2.4 pp. It is used for 17 estimates.

Estimated modelHLE w/ toolsCyberGymSource
Agents-A147.6%76.8%estimated ± 2.4 pp, medium confidence
Apodex 1.156.1%79.2%estimated ± 2.4 pp, medium confidence
Claude Haiku 5.557.4%80.1%estimated ± 2.4 pp, medium confidence
Claude Opus 564.7%89.1%estimated ± 2.4 pp, low confidence
Claude Opus 5.567.7%94.9%estimated ± 2.4 pp, low confidence
Claude Sonnet 557.4%80.1%estimated ± 2.4 pp, medium confidence
Claude Sonnet 5.564.5%88.7%estimated ± 2.4 pp, low confidence
dots3-note Preview52.6%77.8%estimated ± 2.4 pp, medium confidence
GLM-5.3-Flash55.3%78.8%estimated ± 2.4 pp, medium confidence
GPT-6 Astra57.2%79.9%estimated ± 2.4 pp, medium confidence
Nemotron 3 Ultra37.4%76.3%estimated ± 2.4 pp, low confidence
Ornith-1.5-35B-A3B33.4%76.3%estimated ± 2.4 pp, low confidence
Ornith-1.5-397B56.1%79.2%estimated ± 2.4 pp, medium confidence
Ornith-1.5-9B30.5%76.3%estimated ± 2.4 pp, low confidence
Qwen3.7 Max53.5%78.1%estimated ± 2.4 pp, medium confidence
Qwen3.8 Max56.2%79.3%estimated ± 2.4 pp, medium confidence
Step 3.7 Flash47.2%76.8%estimated ± 2.4 pp, medium confidence