OpenAI's o3 solves 4.8% of impossible cases — is this the end of diagnostic dead ends?
openai's o3 deep research model identified new leads for 18 previously unsolved rare disease cases, achieving a 4.8% diagnostic yield in a joint study with boston children's hospital. this demonstrates how reasoning models can cut through complex biomedical data to help experts solve 'undruggable' cases.
researchers from boston children’s hospital and openai published a study in nejm ai demonstrating that the o3 deep research model enabled diagnoses in 18 out of 376 previously unsolved rare disease cases, a 4.8% improvement over specialist analysis alone. the workflow involved feeding the model standardized phenotype terms and genomic variant tables, which it synthesized into evidence-linked explanations for clinicians to validate through standard clinical protocols. unlike typical black-box ai tools, this approach focused on 'explanation-first reasoning' to bridge fragmented medical records and evolving scientific literature, addressing the scalability crisis in genomic reanalysis. this collaboration proves that high-compute reasoning models can practically augment highly specialized medical workflows, suggesting a future where ai periodically re-analyzes stuck cases as new scientific evidence emerges.