Microsoft's MAI-DxO beats generalist doctors on a diagnosis benchmark in a peer-reviewed Science paper

Microsoft AI said on 22 October 2026 that its diagnostic system, MAI-DxO, is now the subject of a peer-reviewed paper in Science. In its announcement, the company says the system reached the right diagnosis far more often than generalist physicians on a step-by-step benchmark, while ordering fewer or cheaper tests.

The system does not give one instant answer. It runs a panel of simulated doctors that propose diagnoses, pick the next question or test, and argue with each other over several rounds. The benchmark, called SDBench, turns 304 New England Journal of Medicine case records and 325 emergency department admissions into interactive cases that start with only a short description of the patient.

Key facts

• Primary: Microsoft AI and the Science paper, 22 October 2026
• Independent: The Verge, 22 October 2026
• Journal cases: 83.2% accuracy with GPT-5, against 36.1% for physicians allowed to use references and 22.3% for physicians without them
• Emergency cases: 80.3% for the system against 46.8% for physicians
• Cost: one budget setting kept 79.6% accuracy at about $1,303 in estimated tests per journal case, compared with $4,725 for physicians using references
• Code: Microsoft is releasing SDBench code under the MIT License; the journal case data needs a license from NEJM Group
• Next step: a prospective study with Beth Israel Deaconess Medical Center in Boston, with no firm start date

How to read it

This is a benchmark result, not a hospital trial. The cases were text only, with no imaging, no messy records and no real patients. The physicians were generalists working alone, and the authors write that the gap should narrow against specialists or care teams. Every author works at Microsoft and holds company stock, and Microsoft built the benchmark it scored best on. Microsoft AI chief Mustafa Suleyman told The Verge the system "needs clinical validation." The open code lets outside groups rerun the tests, and the Boston study is the result to watch.

Independence

Frontier Flux is not affiliated with Microsoft, Science or The Verge. Figures come from the paper, the company post and The Verge's reporting, not from our own testing.