OpenAI says its o3 Deep Research reasoning model was used to help researchers take a second look at 376 previously unsolved childhood rare-disease cases.

The company says the work relied on de-identified clinical and genomic data from those cases. In other words, the model was tested on medical records and genetic information with identifying details removed, while researchers focused on whether the system could surface leads that humans might not have connected before.

After the model’s analysis, expert review and additional testing were used, along with clinical confirmation by physicians. According to OpenAI, diagnoses were ultimately established in 18 of the 376 cases.

This kind of result matters because rare-disease cases can stay unresolved for a long time, even after multiple rounds of testing. When experts are working with the same type of clinical and genomic information, it can be hard to know what pattern to trust first. A reasoning model that supports researchers—rather than replacing doctors—can change the odds of finding the right diagnosis sooner.

Published in NEJM AI

OpenAI also says the study was published on June 18, 2026, in NEJM AI. The article describes how researchers combined the model’s reasoning output with follow-up work and physician confirmation to reach the final diagnoses.

As with any diagnostic tool, this isn’t a claim that a model can diagnose on its own. It’s more about how adding a structured reasoning step can help guide clinicians toward additional tests and clearer answers.

Source: OpenAI