OpenAI’s rare disease research has helped doctors crack 18 previously unsolved childhood cases, after its o3 reasoning model re-analysed genomic data in a study published on Thursday, 18 June 2026.
The work came out of Boston Children’s Hospital, Harvard and OpenAI, and the team pointed the model at 376 cases that specialists had already worked through without landing on an answer, according to a study published in NEJM AI.
The model went back through the data looking for missed leads.
How the OpenAI rare disease model worked
This is where the nerdy detail matters. The o3 Deep Research model did not diagnose anyone. It surfaced evidence-linked hypotheses, candidate explanations tied back to specific genes and symptoms, and handed those to clinicians to test in a lab and confirm.
The humans stayed firmly in charge.
After expert review and follow-up testing, doctors confirmed diagnoses in 18 of the children, an added diagnostic yield of 4.8% on top of what specialists had managed before. In a field where a single answer can end years of uncertainty, that hit rate is a big deal.
There are caveats worth keeping in mind. A 4.8% yield means the model missed the vast majority of the unsolved cases, and every lead still needed human confirmation in a clinical laboratory.
This is assistive technology, not a replacement for the specialists reading the results.
What the 18 rare disease diagnoses covered
The confirmed cases were not minor footnotes. Ten of the children had neurodevelopmental conditions and four had neuromuscular disorders.
Two were cases of early childhood psychosis, and two involved children who had died suddenly, where a diagnosis can carry weight for surviving family members and future siblings.
Rare disease patients often spend years bouncing between specialists in what the field calls a diagnostic odyssey.
The promise here is not a robot doctor, but a tireless second reader that can re-open cold files with the latest gene-disease links and flag connections a tired human might miss.
What the OpenAI study means for AI in medicine
OpenAI has been pushing harder into science and health, and this study lands while the company moves toward a public listing. The careful framing, model suggests and humans confirm, is deliberate.
It keeps clinical responsibility with doctors while showing where large reasoning models can actually pull their weight.
The next step is wider testing beyond a single study, with researchers expected to run the approach across more hospitals and bigger case sets before it becomes a routine tool.
For now, 18 families have answers they did not have last week, which is where the story stands.
