What is HealthBench Professional?
HealthBench Professional is a benchmark published by OpenAI in April 2026 that measures how well language models handle the work clinicians actually bring to them. It contains 525 physician-authored tasks selected from 15,079 real clinician conversations, each graded against a physician-written rubric. It narrows the scope of the original HealthBench, which covered general health conversations, to professional clinical use. The paper is at arxiv.org/abs/2604.27470.
The three use cases
Tasks fall into three categories of clinician work. Care consult: differential diagnosis, management decisions, and treatment questions. Writing and documentation: notes, summaries, letters, and coding. Medical research: synthesizing evidence and answering questions against the literature. A model that is strong in one category can be ordinary in another, which is part of what the benchmark is built to expose.
How the task set was built
| Source pool | 15,079 real conversations that clinicians brought to ChatGPT at work, not synthetic prompts written for the benchmark. |
|---|---|
| Selection | 525 tasks chosen by physicians, with difficult examples deliberately overweighted, roughly 3.5 times their natural rate. |
| Adversarial share | About one third of the set is red-teaming material: tasks written to invite a confident wrong answer. |
| Rubrics | Each task carries its own physician-written rubric. Criteria went through a three-stage construction and adjudication process, with three or more physicians per criterion. |
| Contributors | 190 physicians across 50 countries, 26 specialties, and 52 professional languages. |
How grading works
A model grader (GPT-5.4, low reasoning effort) reads the candidate response and judges each rubric criterion as met or not met. The score is earned points divided by possible points, clipped to 0 to 1, with a length adjustment so a longer answer cannot buy credit by covering everything. Rubrics include negative criteria: a response loses points for harmful or fabricated content regardless of what else it gets right.
Two reference points anchor the scale. Physician-written responses score 0.437 on the same rubrics, and the current model leader scores 0.660. Nothing is near 1.0, and that is by construction: the task set was enriched for difficulty until frontier models fail routinely.
The HealthBench family
| HealthBench May 2025 | 5,000 multi-turn health conversations with laypeople and professionals, graded on 48,562 physician-written rubric criteria from 262 physicians in 60 countries. |
|---|---|
| HealthBench Hard May 2025 | The 1,000 examples frontier models found hardest at release, with top scores around 0.32 at the time. |
| HealthBench Consensus May 2025 | 3,671 examples scored on 34 behavioral criteria validated by physician consensus, built as a near-zero-error baseline. |
| HealthBench Professional April 2026 | 525 tasks narrowed to what clinicians actually bring to a model at work, drawn from real clinician conversations. The benchmark this site tracks. |
The current ranking of 9 models is on the leaderboard. How the numbers on this site are produced is on the methodology page.