Microsoft AI has unveiled MAI-DxO, a diagnostic orchestrator that coordinates multiple AI agents acting like a panel of doctors — questioning patients, ordering tests, and debating diagnoses. A peer-reviewed study published Thursday in Science reports that the system correctly diagnosed about 83% of 304 complex cases drawn from The New England Journal of Medicine, versus about 36% for primary care and general internal medicine physicians with reference materials, and about 22% without. The best individual physician scored 51.8%. Specialists and physician teams who typically handle such complexity were excluded from the comparison.
Microsoft AI CEO Mustafa Suleyman called the outcome a step toward "medical superintelligence," which Microsoft defines as the capacity to surpass the strongest contemporary human-led health systems in medical reasoning and coordination across the full arc of care. Suleyman argued that once a model is in production and improved through real-world usage — the flywheel already seen in coding and chat — progress accelerates quickly, and that medical records and clinical decision support will enter that phase next year.
Microsoft built its own benchmark, SDBench, from published case reports and emergency department records, and plans to open-source it. On cost, GPT-5 alone with a short basic prompt scored about 78.9% on complex cases at an estimated $8,400 per case; MAI-DxO cut cost by roughly $3,000 and raised accuracy by about four percentage points. On emergency department cases the difference was small. The orchestrator is model-agnostic, and on Beth Israel Deaconess cases it reached about 80% accuracy versus about 47% for physicians on average.
Coauthor Dr. Dominic King, Microsoft AI's health lead, outlined five further areas beyond diagnosis — coordinating clinical care, treatment decisions, and the human side of medicine among them. Microsoft is training a medical foundation model with Mayo Clinic, expected next year, and Suleyman anticipates significant gains over the next 24 to 36 months. Hospital partners are preparing clinical evaluations, with real-world testing slated to begin in the coming months.
The study positions AI as decision support rather than a replacement for clinicians. Patients with rare or complex presentations could see fewer missed diagnoses; generalists gain a reasoning and test-ordering assistant that reduces cognitive load; health systems may cut redundant testing and cost. But the hardest cases handled by specialists and teams were not benchmarked, and safety, regulation, and liability in real clinical settings still require independent validation.
Microsoft plans clinical evaluations with hospital partners and will release SDBench. Suleyman expects medical models to improve markedly as companies iterate with real-world deployment. Google and OpenAI have made similar claims about matching or surpassing physicians on diagnostic tasks, intensifying competition. Privacy, bias, and accountability remain open questions as these systems move from simulation to bedside.