Data
Everything the leaderboard shows, every score and every API fact, exists as machine-readable files at URLs that never move. Files and pages regenerate together and share the 2026-09-08 version stamp, and the CC BY 4.0 license means reuse just requires attribution.
Downloads
- the complete results, source document and quote per model
- spreadsheet-ready, one model per row, source name and URL included
- the site indexed for language model agents
- every number on the site in one plain-text file
Citing this leaderboard
A suggested format:
Arcophos. HealthBench Hard Leaderboard. September 8, 2026 snapshot. https://healthbenchhard.ai. Accessed [date].
The benchmark itself is cited via OpenAI's paper at arxiv.org/abs/2505.08775. Numbers here are compiled from published documents, each named on the sources page and carried in the files as a source object per model; when citing a single score, cite its source document as well.
Versions
| 2026-09-08 | Current file, 16 models. Adds a source object, confidence, and reporting fields per model and a top-level sources list. |
| 2026-08-16 | First release, scores only. |
Download URLs are permanent. Fields get added, never renamed. Every change is announced on the updates page.