Data
The complete leaderboard, every score, its source document, and every API fact on this site, is published in machine-readable form at stable URLs. The files carry the same 2026-09-30 version stamp as the pages, are regenerated together with them, and are licensed CC BY 4.0, so they can be reused with attribution.
Downloads
- full results with benchmark metadata
- one row per model, spreadsheet-ready
- site index for language model agents
- the whole leaderboard as one plain-text document
Citing this leaderboard
Suggested format:
Arcophos. HealthBench Professional Leaderboard. September 30, 2026 snapshot. https://healthbenchprofessional.com. Accessed [date].
For the benchmark itself, cite OpenAI's paper: arxiv.org/abs/2604.27470. The scores here are compiled from published documents, each listed on the sources page, and a number read from one document can differ from the same model's number in another for the reasons given on the methodology page. The JSON carries a source object per row (document, URL, quote, locator, confidence) and the CSV carries source_name and source_url columns.
Versions
| 2026-09-30 | Current file: 31 models, exported from the results database with a source field per row. |
| 2026-08-16 | Initial release: 31 models, no source fields. |
The download URLs never change. New fields may be added; existing fields are not renamed. Changes are announced on the updates page.