OpenAI’s reasoning model found 18 new rare disease diagnoses where doctors were stuck

I’ve been following the intersection of AI and medicine for a while now, and this one actually surprised me. OpenAI published a paper yesterday showing that one of their reasoning models helped diagnose rare genetic diseases in children—18 new diagnoses in cases where doctors had already exhausted every conventional option.

That’s not a small number. These are kids who’ve been through the diagnostic wringer: multiple specialists, genetic tests that came back inconclusive, families bouncing from hospital to hospital. The model didn’t just match symptoms to known diseases. It reasoned through ambiguous genetic variants and clinical presentations that humans had missed or couldn’t resolve.

The approach is straightforward but clever. They fed the model structured data from electronic health records—lab results, clinical notes, family history, previous genetic sequencing reports—and asked it to generate differential diagnoses ranked by likelihood. Then they had physicians review the top candidates. In 18 cases, the model’s suggestions led to confirmed diagnoses that had previously eluded every clinician involved.

What’s interesting is that this isn’t a “look up a database” trick. Rare diseases are rare precisely because they don’t show up in standard diagnostic algorithms. The model had to reason about novel combinations of symptoms and genetic markers. That’s where the “reasoning” part of the model name matters.

I’ve seen plenty of AI-in-medicine hype that fizzles once you look at the actual numbers. But 18 new diagnoses in a single study, from cases that were considered solved or unsolvable—that’s higher than I expected. It suggests we’re moving past the phase where AI just flags obvious things doctors already know.

The catch? This still requires a human in the loop. The model generates candidates, but a physician has to validate each one. That’s fine for now. I’d rather have a cautious, collaborative system than an autonomous black box making calls about kids’ health.

OpenAI didn’t release the model or the specific methodology publicly, which is frustrating but understandable. Medical applications of AI carry real liability. Still, I’d love to see independent replication of these results. One study, even a strong one, isn’t a revolution.

For now, this is a solid proof of concept. If you’re a parent with a child who has an undiagnosed rare disease, this probably doesn’t change your situation tomorrow. But it does mean researchers have another tool in the toolbox. And for the families of those 18 kids, it changes everything.