On October 22, 2026, Science published Microsoft's study of MAI-DxO, an orchestrator that coordinates several AI agents acting like a panel of doctors. They question a patient, order tests, and debate a diagnosis. On 304 hard case records from the New England Journal of Medicine, MAI-DxO scored 83%. Generalist doctors scored 36% with reference materials and 22% without. The paper is peer-reviewed, but it is Microsoft's own study, and the test was a simulation.
What the Numbers Show
The doctors were primary care physicians and general internists, each working alone. TechCrunch notes that cases this complex are usually handled by specialists or teams, who were not in the comparison. Each record was turned into a step-by-step exercise where the doctor or the AI asks for information and orders tests, and each test carries a virtual cost. No real patients were involved.
On more typical emergency department cases from a public database called MIMIC-IV-ED, Microsoft reports 80% for MAI-DxO against 47% for physicians. Microsoft also says MAI-DxO reached answers at about half or a third of the cost, and that no other human or AI approach was both more accurate and cheaper.
What Beat the Doctors
The study's own data suggests the AI model underneath matters a great deal. According to TechCrunch's reading of Microsoft's data, GPT-5 on its own, with a short prompt, scored 78.9% on the complex cases at about $8,400 per case. MAI-DxO cut that cost by about $3,000 but added only about four points.
The newer GPT-5.5 went from 72% to nearly 87% with MAI-DxO on complex cases. On emergency cases, MAI-DxO added little: GPT-5.5 scored about 85% with or without it. That is close to the 80% Microsoft reports for MAI-DxO there, though the underlying models may differ. Gemini 2.5 Flash rose from 46.4% to nearly 69%. TechCrunch reports that the gap between weaker and stronger models was larger than MAI-DxO's boost. A Microsoft spokesperson said gains naturally shrink as models near the maximum, and Microsoft says the gains held with newer models, so the approach is not tied to one model.
So much of the 83% reflects a strong model. The orchestrator helps most on complex cases and with weaker models.
What the Test Did Not Cover
The cases were written records from the past. Each doctor worked alone, without a team. The test covered diagnosis only, not treatment. Microsoft has open-sourced the benchmark for select NEJM cases, so others can try to reproduce the results, but no independent replication was found. The headline number is also not new: a June 2025 Microsoft preprint reported 85.5% with OpenAI's o3 model on the same 304 cases, against 20% for 21 physicians.
What Could Come Next
Microsoft is partnering with Beth Israel Deaconess Medical Center in Boston. Patients will get standard care, and those who consent will also use MAI-DxO, so its advice can be compared with the doctors'. According to TechCrunch, Google published a real-world urgent care study of its AMIE system with Beth Israel in The Lancet this month, so the race to real-world evidence has started. Microsoft AI chief Mustafa Suleyman calls this work a step toward "medical superintelligence," his term, and says Microsoft is working with Mayo Clinic on a healthcare model planned for next year. The long-term picture is speculation. He predicts healthcare models will improve by an order of magnitude within 24 to 36 months. That is a forecast, not a result.
Disclosure: This article was drafted with Claude, an AI model made by Anthropic. Anthropic models were among those tested in Microsoft's earlier work and are named among the companies whose models Microsoft tested. Claims were checked against Microsoft's posts and outside reporting, and Microsoft's own statements are labeled as such.
