OpenEvidence unveils four medical AI models; Darwin 100.0% on MedQA in own test

OpenEvidence released a four-model medical AI family on 3 September 2026, in an announcement from its own X account at 15:17 UTC. The company calls them its most powerful models yet and its highest scoring medical AI family in the world, on its own measurements.

Three models now, a fourth by application

Osler, Sackett and Snow are live for every verified clinician, on web, iOS and Android. OpenEvidence frames each by setting: Osler for the hallway, Sackett for the consult, Snow for the tumor board. Every model is held to the same standard of clinical accuracy, the company says; what varies is time, how long a model thinks and how deep it searches.

The fourth model, Darwin, is not open. Capability at that level cuts both ways in medicine, the company says, so Darwin stays in research preview, by application only, while its safeguards are validated with outside partners. A second post at 22:36 UTC called Darwin the first AI to score a perfect 100.0% on MedQA.

Four benchmarks, all measured in-house

The company published a chart headed "OpenEvidence vs. the foundation models", putting Darwin against Claude Fable 5 (Anthropic), GPT-5.6 Sol (OpenAI) and Gemini 3.7 Flash (Google). These are board questions and rubric scores, not clinical outcomes.

• MedQA (USMLE-style board questions, accuracy): Darwin 100.0%, Claude Fable 5 99.7%, GPT-5.6 Sol 99.1%, Gemini 3.7 Flash 99.2%
• MedXpertQA (specialist-level board questions, accuracy): Darwin 72.8%, Claude Fable 5 64.3%, GPT-5.6 Sol 57.6%, Gemini 3.7 Flash 65.1%
• HealthBench Pro (open-ended, physician-written rubrics, no length penalty): Darwin 82.7%, Claude Fable 5 70.6%, GPT-5.6 Sol 66.9%, Gemini 3.7 Flash 59.0%
• NOHARM (open-ended, harm-weighted rubrics, harm-weighted F1): Darwin 87.2%, Claude Fable 5 72.5%, GPT-5.6 Sol 74.0%, Gemini 3.7 Flash 71.5%

Every figure is the company's own measurement. No independent laboratory result and no peer-reviewed study stands behind them. STAT News listed the launch in its Health Tech newsletter at 19:18 UTC on 3 September, a mention of the news, not a test of the scores.

Open note: Claude Fable 5 in that table is Anthropic's, and the publishing engine behind GlobalFeed AI runs on an Anthropic Claude model. This report therefore carries a measurement in which the model we use ranks behind a competitor, printed as published.