Microsoft MAI-DxO Beats Doctors 83% to 36% in Science Study
3 min readMicrosoft says its diagnostic AI system, MAI-DxO, correctly solved 83 percent of complex medical cases, while general physicians working alone managed 36 percent. The results were published Thursday in the journal Science, and Microsoft AI chief Mustafa Suleyman called them a step toward what he describes as medical superintelligence.
What Microsoft MAI-DxO Is
MAI-DxO, short for the Microsoft AI Diagnostic Orchestrator, is not a single model. It coordinates several AI agents that behave like a panel of doctors. The agents question a simulated patient, order tests, weigh the cost of each test, and debate a diagnosis before committing to an answer. Microsoft first previewed the system in mid-2025; the Science paper is its first peer-reviewed showing.
The evaluation used complex cases from the New England Journal of Medicine, the kind of puzzles usually handled by specialists or physician teams. Microsoft ran the orchestrator on models from OpenAI, Google, Anthropic, Meta, and DeepSeek, with the best results coming from OpenAI’s GPT-5 and GPT-5.5.
What the Numbers Show
According to TechCrunch, the 83 percent figure is measured against primary care physicians and general internists working alone with reference materials. Specialists were not part of the comparison, and the cases were simulations rather than live patients.
The paper’s own cost data complicates the headline. GPT-5 on a short prompt scored 78.9 percent at about $8,400 per case; adding MAI-DxO trimmed roughly $3,000 from that bill but lifted accuracy by only about four points. GPT-5.5 gained more, climbing from about 72 percent alone to nearly 87 percent with the orchestrator. On emergency department cases, the system added almost nothing for either model. Microsoft’s researchers acknowledge that most of the gain comes from the underlying models, with orchestration adding a layer on top.
Why It Matters
Microsoft now joins Google and OpenAI in claiming its AI can match or beat doctors on hard cases. Dominic King, vice president of health at Microsoft AI, framed the problem as a widening gap between demand for medical expertise and the supply of it. Google published its own study this month in The Lancet, testing its AMIE system in real urgent care visits.
The next step moves out of simulation. Microsoft will partner with Beth Israel Deaconess Medical Center in Boston to test MAI-DxO in real hospital settings, where consenting patients still receive standard care while the system’s recommendations are compared against those of human doctors. Microsoft is also working with Mayo Clinic on a healthcare foundation model slated for next year, and Suleyman expects health models to improve dramatically within 24 to 36 months.
Benchmarks on textbook cases are one thing; a hospital ward is another. The Beth Israel trial will show whether an AI panel of doctors holds up when the patient is real.