Microsoft's MAI-DxO, a diagnostic orchestrator that runs GPT-5 and GPT-5.5 underneath, correctly diagnosed 83% of complex New England Journal of Medicine case records against 36% for generalist physicians working the same set. The result is published in Science, and clinical testing at Beth Israel Deaconess is expected within two months.
The caveats are worth stating alongside the number, because the headline ratio does a lot of work. NEJM case records are written-up diagnostic puzzles with the relevant findings already assembled - they are not patients presenting in a clinic. The physician comparison group were generalists working without access to colleagues, textbooks or search. And no real-world trial has run yet, which is exactly what the Beth Israel work is for.
What the study does establish is that an orchestration layer over existing models beats a single model prompt on sequential diagnostic reasoning. That finding generalises further than the 83% does.
Sources
WebProNews · Technology.org · The Neuron