HealthBench & HealthBench Professional

artificialanalysis.aiDR 85 Accessible 300 views

Lectura y aprendizaje Artificial Analysis leaderboards for OpenAI's HealthBench and HealthBench Professional benchmarks, which grade AI model responses to realistic health conversations against physician-created rubrics.

Visitar ahora Guardar en favoritos

Editor's Review

For reference only

HealthBench & HealthBench Professional page preview

The value here lies in separating everyday-user health chats from clinician-facing tasks, with the Professional set adding adversarial and difficult conversations. Reporting both raw rubric scores and response length exposes how verbosity can inflate results. Useful for model selection in health-adj

Site introduction

Artificial Analysis has added leaderboards for two OpenAI benchmarks that assess how AI models handle realistic health conversations. Rather than measuring medical knowledge in isolation, the benchmarks grade actual model replies against rubrics written by physicians, making the results closer to what users would encounter in practice.

HealthBench draws on 5,000 conversations with everyday users and health workers across seven themes, ranging from emergency referrals to global health. Each rubric criterion maps to one of five axes: accuracy, completeness, communication quality, context awareness and instruction following. HealthBench Professional is a smaller set of 525 conversations, all with clinicians, covering consultation, writing and research tasks, and includes adversarial and difficult exchanges designed to challenge models.

The leaderboards report both the raw score, meaning the share of rubric points earned, and response length. Showing length alongside score matters because longer answers can accumulate rubric points, so readers can judge whether a high score reflects substance or verbosity.

These leaderboards suit teams building or procuring health-related AI features, researchers comparing model behavior on clinical communication, and anyone who wants a rubric-based view of model performance beyond generic benchmarks. When using the results, check both score and length, and note that HealthBench and Professional measure different audiences and should not be tr