In brief
The system reached the correct diagnosis in 83.2% of 304 NEJM case reports, against 36.1% for generalist doctors with textbooks and the internet and 22.3% for those without them.
On 325 emergency department admissions, the system scored 80.3% compared with 46.8% for human doctors.
The doctors were generalists, the cases were text only, and neither side had physical examinations or medical imaging.
Microsoft's team plans to test with Beth Israel Deaconess Medical Center in Boston whether the tool can help triage patients.
How the system works
Microsoft AI has built the MAI-DxO, a system that works through a patient case in stages rather than reading every symptom, lab value and scan at once. It first decides which history questions to ask and which parts of the examination matter, then selects the tests to order. As results come back, it revises its working diagnosis and requests further tests where needed. A coordinating layer, called the orchestrator, runs several simulated clinical experts who propose competing explanations and check them against each other. The study was published on Thursday, 22 October 2026, in the journal Science.
Accuracy against doctors
In tests on 304 complex case reports from The New England Journal of Medicine, the system reached the correct diagnosis in 83.2% of cases. Generalist physicians who could consult textbooks and the internet scored 36.1%, while those working without such references scored 22.3%.
- MAI-DxO
- 83.2 %
- Generalist doctors with references
- 36.1 %
- Generalist doctors without references
- 22.3 %
Data from the article text
In a separate test on 325 emergency department admissions, the system scored 80.3% against 46.8% for the doctors. The Verge reported Microsoft's claim that the tool can diagnose conditions up to four times more accurately than human doctors, although the paper does not include newer AI models in its comparison.
- MAI-DxO
- 80.3 %
- Human doctors
- 46.8 %
Data from the article text
What the researchers say
Microsoft's researchers also measured how well the system chose tests, not only whether it reached the right answer. Harsha Nori, head of AI at Microsoft AI Health, put the team's aim in these terms.
“We're not just measuring the accuracy of the systems.”
Dominic King, a corporate vice president in Microsoft AI's health unit, argued that most medical AI has been graded on multiple-choice questions. He said such systems have learned to score well on those tests, and that this does not reflect how patients present in practice.
“No patient ever came to see me with a nice summary.”
Mustafa Suleyman, Microsoft AI's CEO, called the result a step toward "medical superintelligence". Spider's Web reported that he believes AI could match doctors within a few years.
Cost and model choice
On complex cases, GPT-5 working alone scored 78.9% at about $8,400 per case, according to TechCrunch. Adding the orchestrator cut that cost by about $3,000 and lifted accuracy by only about four points. With GPT-5.5, a model released in April 2026, accuracy on complex cases rose from 72% to nearly 87% at a slightly lower cost. Gemini 2.5 Flash scored 46.4% alone and nearly 69% with the orchestrator. The cost figures are estimates based on the cases rather than actual medical bills.
Limits and next steps
The comparison has clear limits. The doctors were generalists rather than specialists, the cases were presented as text, and neither the AI nor the doctors had access to physical examinations or medical imaging. An earlier preprint from June 2025 reported gains with models from Google, Anthropic and Meta. Microsoft's team plans to test with Beth Israel Deaconess Medical Center in Boston whether the tool can help sort patients by urgency. The company acknowledges that a long road remains before the technology can be widely deployed.