Microsoft AI's diagnostic research has passed peer review. A paper published on October 22 in Science describes MAI-DxO, the company's Diagnostic Orchestrator, and reports that on retrospective cases it reached the correct diagnosis far more often than generalist physicians did. Microsoft AI chief executive Mustafa Suleyman presented the result as a step toward what the company calls medical superintelligence, while conceding that the system still needs clinical validation.
MAI-DxO is not a single model. It coordinates several AI agents that behave like a small panel of clinicians: they ask the simulated patient questions, decide which tests to order, argue over possible explanations and then commit to a diagnosis. The study scores both accuracy and the estimated cost of the tests requested. Microsoft built its own benchmark for this, SDBench, from clinicopathological case records that appeared in The New England Journal of Medicine and from emergency-department admissions in a large database from Beth Israel Deaconess Medical Center in Boston, and The Verge reports that the company plans to open-source it.
On the complex NEJM cases, the system landed on the right answer in roughly 83 percent of attempts. Generalist physicians, meaning primary care doctors and internists, working alone reached about 36 percent with reference materials and 22 percent without them. On emergency-department cases MAI-DxO scored about 80 percent against an average of roughly 47 percent for physicians, and Microsoft says it tended to get there with fewer and cheaper tests. TechCrunch points out an important limit of the comparison: cases this hard are normally handled by specialists or teams, who were not part of the test.
The paper also shows how much depends on the underlying model. The published configuration ran on OpenAI's GPT-5, and Microsoft tested models from OpenAI, Google, Anthropic, Meta and DeepSeek. According to figures reported by TechCrunch, GPT-5 with a basic prompt already reached 78.9 percent on the complex cases at an estimated cost of around $8,400 per case; the orchestrator lowered that cost by about $3,000 while adding roughly four points of accuracy. With GPT-5.5 the orchestrator lifted complex-case accuracy from 72 percent to nearly 87 percent, yet brought little benefit on emergency cases, where both setups landed near 85 percent. The weakest models gained most: Gemini 2.5 Flash rose from 46.4 to nearly 69 percent.
The next step moves from archives to patients. Microsoft says a newly agreed collaboration with Beth Israel Deaconess will let it run a prospective study in which patients interact with the system directly and its conclusions are compared with physicians' independent assessments across hospital settings. Suleyman also told TechCrunch that Microsoft is working with the Mayo Clinic on a healthcare foundation model it hopes to release next year.
For clinicians and hospital leaders, the paper offers a structured way to think about AI as a diagnostic second opinion that is measured on cost as well as correctness. For AI developers, it is evidence that orchestration can add value on hard reasoning problems, though the gains shrink as base models improve.
Several caveats travel with the headline numbers. Every author is a Microsoft employee, the benchmark is Microsoft's own, the comparison excludes specialists, and all cases were historical. The paper does not cover the newest frontier models. Until the prospective trial reports, claims about real-world clinical performance remain untested.