HealthBench Professional Leaderboard
Last updated August 16, 2026 · 9 models evaluated
Claude Fable 5 leads the HealthBench Professional leaderboard at 0.660, ahead of GPT-5.6 Sol (0.605) and Claude Opus 5 (0.598). The benchmark, published by OpenAI in April 2026, scores models on 525 tasks physicians selected from real clinician conversations, graded against physician-written rubrics on a 0 to 1 scale. Physician-written responses score 0.437 on the same rubrics.
Leaderboard
525-task set, one response per task. Updated August 16, 2026.
Full ranking
| # | model | score | size | context | cost in / out per 1M | license | |
|---|---|---|---|---|---|---|---|
| 1 | | Claude Fable 5 Anthropic | 0.660 | — | 1.0M | $10.00 / $50.00 | proprietary |
| 2 | | GPT-5.6 Sol OpenAI | 0.605 | — | 1.1M | $5.00 / $30.00 | proprietary |
| 3 | | Claude Opus 5 Anthropic | 0.598 | — | 1.0M | $5.00 / $25.00 | proprietary |
| 4 | | Claude Sonnet 5 Anthropic | 0.578 | — | 1.0M | $2.00 / $10.00 | proprietary |
| 5 | | GPT-5.6 Terra OpenAI | 0.577 | — | 1.1M | $2.00 / $12.00 | proprietary |
| 6 | | Claude Opus 4.8 Anthropic | 0.558 | — | 1.0M | $5.00 / $25.00 | proprietary |
| 7 | | GPT-5.6 Luna OpenAI | 0.557 | — | 1.1M | $0.20 / $1.20 | proprietary |
| 8 | | GPT-5.5 Instant OpenAI | 0.384 | — | 400K | $5.00 / $30.00 | proprietary |
| 9 | | MAI-Thinking-1 Microsoft | 0.350 | 1.0T | — | — | proprietary |
Rubric credit on the 525-task set, graded with the benchmark's default grader (GPT-5.4, low reasoning effort), one response per task. List API prices per 1M tokens as of August 16, 2026. GPT-5.6 models charge higher rates above 272K input tokens. MAI-Thinking-1 is in public preview on Microsoft Foundry without final list pricing. Each model name links to its full result.
What the benchmark measures
HealthBench Professional evaluates language models on the work clinicians actually bring to a model: care consults such as differential diagnosis and management questions, clinical writing and documentation, and medical research. The 525 tasks were selected by physicians from 15,079 real clinician conversations, difficult cases were deliberately overweighted, and roughly a third of the set is adversarial. 190 physicians across 50 countries and 26 specialties wrote and adjudicated the rubrics. The full construction is described on the benchmark page.
How to read the scores
Scores are earned rubric points divided by possible points, clipped to 0 to 1, with a length adjustment that keeps verbosity from buying credit. Absolute numbers run low by design: the field's best result is 0.660, and physician-written responses score 0.437 on the same rubrics. Scores depend on the grader model, so numbers graded under a different configuration are not directly comparable; every score on this site comes from the same run protocol, described on the methodology page.
Head to head
Score, price, and context differences for the pairings people actually weigh, with the full score-difference matrix on the compare page.
- 0.660 vs 0.605 · Claude Fable 5 by 0.055
- 0.605 vs 0.598 · GPT-5.6 Sol by 0.007
- 0.578 vs 0.577 · Claude Sonnet 5 by 0.001
- 0.660 vs 0.598 · Claude Fable 5 by 0.062
- 0.598 vs 0.558 · Claude Opus 5 by 0.040
- 0.557 vs 0.384 · GPT-5.6 Luna by 0.173
Data
The full results are published as JSON and CSV at stable URLs, licensed CC BY 4.0. Citation format and version history are on the data page.