Microsoft says its MAI-DxO diagnoses hard cases more than twice as accurately as doctors
The cases came from medical journals and hospital records, and another model played the part of the patient's chart.
On October 22, 2026, Microsoft AI published a post called Building Medical Superintelligence, with the research behind it out the same day in Science. Its diagnostic system, MAI-DxO, was more than twice as accurate as physicians using textbooks and the internet on the hardest cases from The New England Journal of Medicine.
Without those resources, the doctors fell further behind: Microsoft puts MAI-DxO at nearly four times as accurate. It says the system got there at half or even one third of the cost, with fewer tests and procedures.
- System
- MAI-DxO, our AI diagnostic system
- Emergency cases
- 1.7 times more accurate than physicians, on cases from MIMIC-IV-ED
- NEJM cases
- more than twice as accurate as physicians using textbooks and the internet
- Cost
- half or even one third of the cost
- Published
- Science, with a vision paper in BMJ Digital Health and AI
How the test works
Microsoft calls the test SDBench, for Sequential Diagnosis Benchmark, and its harness is on GitHub. There's no multiple choice. A system gets a brief case summary that doesn't give the diagnosis away, then has to work the case up turn by turn, the way a doctor would.
Each turn, it can order a test, ask a question or commit to a diagnosis. Another model, the gatekeeper, answers from the full case write-up, and makes up a realistic result when the write-up doesn't contain one. When the system commits, a judge scores it against the real diagnosis from 1 to 5, and a 4 or 5 counts as correct (the judge is a model too, so every score is one model grading another).
Each case is scored on both diagnostic accuracy and the cost of the tests ordered
That second score is why the cost figures matter. A model that orders every scan in the building would rack up a big bill, and the benchmark charges it for that.
Where do the cases come from? Two places. For the hard ones, Microsoft used recent NEJM Clinical Pathological Conference cases, published as Case Records of the Massachusetts General Hospital, and you have to license them from NEJM to run them yourself. Emergency cases come from de-identified hospital records on PhysioNet, closer in form to discharge summaries. Out of the box, the public repository runs five synthetic cases with GPT-5.
Mustafa Suleyman, Microsoft AI's CEO, is on the Science paper's author list, along with Eric Horvitz.
What it means for patients
The emergency department is where Microsoft's 1.7 times figure lands, and it's where the US government has tried to count mistakes. A December 2022 review for the Agency for Healthcare Research and Quality estimated that 1 in 18 emergency patients receive an incorrect diagnosis.
which translates to 7.4 million patients misdiagnosed every year (or 5.7% of all ED visits annually)
Five conditions came out most vulnerable to misdiagnosis: stroke, heart attack, aortic aneurysm or dissection, spinal cord injury and blood clots. Women and non-White patients were often at higher risk. Not everyone accepts these numbers. AHRQ's own patient safety site lists a 2023 response titled "the wrong answer to the wrong question", and the report was updated on August 14, 2023 with errata and an addendum.
So the gap Microsoft is aiming at is real, even if its size is disputed. None of the new numbers come from live patients, though. Every result in the post comes from cases that already happened and were written up, with a model playing the patient's chart.
Google's latest test had patients in the room
On October 8, 2026, Google described a study in The Lancet in which 98 patients at Beth Israel Deaconess Medical Center's primary care clinic talked to AMIE, its diagnostic chatbot, before urgent care visits. Supervising physicians watched in real time, and none of the conversations had to be stopped under the study's safety rules. AMIE's differential diagnoses matched the doctors' final diagnoses 90% of the time.
I'd read the two side by side. Microsoft has the bigger accuracy claim on harder cases, and Google has the live patients. Google itself says larger clinical trials are needed before patient-facing AI can be judged at scale.
Microsoft's vision paper on medical superintelligence came out the same week in BMJ Digital Health and AI.