Microsoft’s AI Diagnosed 83% of Hard Cases; Doctors Got 36%

Microsoft’s MAI-DxO system reached the right diagnosis in 83% of complex cases from the New England Journal of Medicine, while primary care doctors and general internists using reference materials got 36% right, according to a study in Science that Microsoft announced on Oct 22, 2026. Without textbooks or the internet, the doctors’ accuracy fell to 22%.

The cases were simulated. MAI-DxO, which runs several AI agents as a virtual panel of doctors, and the physicians each began with limited information and chose which questions to ask and which tests to order, with every test carrying an estimated cost. The benchmark draws on 304 NEJM case records.

On more routine emergency department cases, taken from a US hospital dataset called MIMIC-IV-ED, MAI-DxO scored 80% against 47% for physicians, Microsoft says.

The comparison has limits. Cases this complex usually go to specialists or teams of doctors, and neither was tested, TechCrunch noted. The study’s own data also suggests much of the advantage comes from the underlying model. OpenAI’s GPT-5 on its own, with a basic prompt, scored 78.9% on the complex cases at about $8,400 per case; MAI-DxO’s orchestration raised that by about four points and cut roughly $3,000. On emergency cases, the orchestration added little.

The paper is the peer-reviewed version of work Microsoft first released as a preprint in June 2025. That version reported up to 85.5% accuracy for the system, against an average of 20% for 21 US and UK physicians. Microsoft AI chief Mustafa Suleyman called the new result a step toward “medical superintelligence,” TechCrunch reported.

MAI-DxO is not a product doctors can use. Microsoft says the real test is performance in clinical settings and is starting a study with Beth Israel Deaconess Medical Center in Boston, where patients who agree will use the system and its conclusions will be compared with their physicians’ own. It is also releasing parts of its benchmark, SDBench, so other researchers can rerun the test.