benchgap
Calibration

GPQA → HealthBench Professional

HealthBench Professional is estimated from GPQA with a Hill curve fitted on 7 models measured on both: y = 0.0000 + (1.2000 − 0.0000)·x^6.00 / (0.94581^6.00 + x^6.00), R² = 0.65, cross-validated error 4.2 pp. It is used for 15 estimates.

Estimated modelGPQAHealthBench ProfessionalSource
GPT-5.292.4%55.8%estimated ± 4.2 pp, medium confidence
Qwen3.7 Max92.4%55.8%estimated ± 4.2 pp, medium confidence
Gemini 3.5 Flash92.2%55.4%estimated ± 4.2 pp, low confidence
Qwen3.8-Omni-Flash91.0%53.1%estimated ± 4.2 pp, low confidence
DeepSeek V4.1 Flash 101790.9%52.9%estimated ± 4.2 pp, low confidence
Kimi K2.690.5%52.1%estimated ± 4.2 pp, low confidence
Qwen3.6 Plus90.4%51.9%estimated ± 4.2 pp, low confidence
Qwen3.7 Plus90.3%51.7%estimated ± 4.2 pp, low confidence
DeepSeek V4 Pro90.1%51.3%estimated ± 4.2 pp, low confidence
Grok 4.390.1%51.3%estimated ± 4.2 pp, low confidence
Claude Sonnet 4.689.9%50.9%estimated ± 4.2 pp, low confidence
Interfaze Beta89.9%50.9%estimated ± 4.2 pp, low confidence
Qwen3.5 397B88.4%48.0%estimated ± 4.2 pp, low confidence
DeepSeek V4 Flash 073188.1%47.4%estimated ± 4.2 pp, low confidence
Qwen3.6-27B87.8%46.8%estimated ± 4.2 pp, low confidence

Fitted on

ModelGPQAHealthBench Professional
GPT-5.492.8%48.1%
GPT-5.6 Luna92.3%55.7%
GPT-5.6 Sol94.6%60.5%
GPT-5.6 Terra92.9%57.7%
GPT-6.2 Astra97.0%66.3%
GPT-6.2 Sol96.0%65.9%
GPT-6 Astra96.0%64.7%