Microsoft's AI "doctor panel" beat physicians on hard diagnoses, and now it's in Science

Microsoft AI said on Thursday, Oct. 22, 2026, that peer-reviewed results for its MAI-DxO diagnostic orchestrator were published in Science, showing it outperformed generalist physicians on complex and emergency-department cases at lower cost.

Getting a headline number into Science is a real step up from a company blog post, and 83% versus 36% is the kind of gap that gets attention. But read the fine print: the doctors were generalists, not specialists, and a big chunk of the accuracy comes from the frontier model underneath, with the orchestration adding a few points and cutting the bill. The more interesting result may be cost: an AI that orders fewer pointless tests is something hospitals would actually pay for. Until regulators and real clinics get their hands on it, "medical superintelligence" is still a slogan with a very good spreadsheet.

MAI-DxO is the Microsoft AI Diagnostic Orchestrator. It coordinates several AI agents that act like a panel of doctors. They question the "patient," order tests, and debate a diagnosis step by step, and they do it against a cost budget, a limit on how much testing the panel is allowed to spend. The patient, here, is the case they are working, not a person in a clinic. Those lines are Microsoft's, from its Oct. 22 account of results published the same day in Science.

On complex case challenges from the New England Journal of Medicine, MAI-DxO was correct 83% of the time. Generalist physicians, primary care doctors and general internists, got 36% when they could use reference materials, the books and databases a doctor can look things up in. Microsoft says the gap is nearly four times when those physicians have no resources. Specialists were not part of the comparison. Those figures are Microsoft's.

On emergency-department cases from MIMIC-IV, a public research archive of hospital visits, Microsoft says MAI-DxO was 1.7 times more accurate than physicians, at one-half to one-third of the cost. Those lines are Microsoft's.

A lot of the score is the model underneath the panel. GPT-5 alone got 78.9% on the hard cases, at about $8,400 per case. With the orchestrator, the bill was about $3,000 lower and the score about 4 points higher. GPT-5.5 went from 72% to nearly 87%. Gemini 2.5 Flash went from 46.4% to about 69%. On the emergency-department cases there was little difference, about 85% with the orchestrator or without. Those figures are Microsoft's account of the work.

Mustafa Suleyman, chief executive of Microsoft AI, calls the result a step toward "medical superintelligence." Microsoft is working with Mayo Clinic on a healthcare foundation model planned for next year. A foundation model is a large model trained broadly, then adapted to a field. Dominic King, vice president of health at Microsoft AI, said, "There is something very substantial and exciting here." Those lines are theirs.

MAI-DxO is not approved for clinical use, and it is not available for clinical use. The scores come from journal case challenges and a research archive. They are not a report of the system running in a hospital.

The picture is Microsoft's own press photo of Mustafa Suleyman speaking on stage at a Microsoft event in April 2025. Dan DeLong took the photograph.

In plain terms, Microsoft said on Thursday that a panel of AI agents, with results now in Science, got 83% of hard journal cases right, against 36% for generalist doctors who could look things up. Much of that score is the model underneath. The panel adds a few points and a smaller bill. It is not cleared for clinical use.