HealthBench
HealthBench is a medical-domain AI benchmark released by OpenAI, designed to systematically evaluate AI models' capabilities in healthcare scenarios, including question answering, diagnostic suggestions, and clinical reasoning. The benchmark covers multiple medical professional scenarios and employs a rigorous scoring mechanism to measure the accuracy and safety of model outputs, providing a standardized evaluation framework for the research and comparison of medical AI systems.
Timeline (last 90 days)
HealthBench is a public evaluation benchmark released by OpenAI in May 2025, comprising roughly 5,000 multi-turn conversational evaluation items written by practicing physicians from multiple countries, each paired with physician scoring criteria.
HealthBench Professional is a more difficult subset of the HealthBench system aimed at professional clinical scenarios, with dialogues closer to real physician consultations or specialty referrals.
AntAngelMed is the best-performing open-source model in the OpenAI HealthBench evaluation, with a particularly clear advantage on the Hard subset.
In HealthBench, doctors first wrote the grading rubric, and OpenAI's GPT-4.1 evaluator exceeded the average doctor's level on 5 of 7 topics.
All Facts (6)
此前的HealthBench健康基准测试存在被游戏化的问题：模型给出的答案越长，得分就越高
85%Unverified各大AI实验室发现HealthBench的冗长加成漏洞后，纷纷利用这种机制来刷分
75%UnverifiedHealthBench Professional is a more difficult subset of the HealthBench system aimed at professional clinical scenarios, with dialogues closer to real physician consultations or specialty referrals.
50%UnverifiedAntAngelMed is the best-performing open-source model in the OpenAI HealthBench evaluation, with a particularly clear advantage on the Hard subset.
50%UnverifiedIn HealthBench, doctors first wrote the grading rubric, and OpenAI's GPT-4.1 evaluator exceeded the average doctor's level on 5 of 7 topics.
50%UnverifiedHealthBench is a public evaluation benchmark released by OpenAI in May 2025, comprising roughly 5,000 multi-turn conversational evaluation items written by practicing physicians from multiple countries, each paired with physician scoring criteria.
50%