Microsoft’s MAI-DxO scores 83% on simulated complex cases

Microsoft says its Diagnostic Orchestrator (DxO) achieved higher diagnostic accuracy than generalist physicians across two sets of medical cases, with the system scoring about 80% on emergency department cases and 83.2% on clinicopathological cases from The New England Journal of Medicine.

The reported results come from benchmark testing, not documented use in routine patient care. On the New England Journal of Medicine cases, physicians averaged 22.3% accuracy without reference tools and 36.1% when they could use them, according to the study figures reported by medicaleconomics.com.

On emergency department admissions, Microsoft’s system recorded 80.3% accuracy, compared with 46.8% for physicians. The Verge gives rounded figures of 80% and 47%, respectively, for that comparison. The emergency department cases came from a large patient-admissions database at Beth Israel Deaconess Medical Center in Boston.

The Verge reports that the system also beat doctors on this set. Medicaleconomics.com emphasizes a limitation of the comparison: it describes the benchmark as controlled and text-only, and cautions against treating the results as proof of performance in real-world practice.

The Verge reports that DxO uses OpenAI’s GPT-5 and that the researchers found it outperformed frontier AI models operating independently. The system organizes multiple AI agents to work through diagnostic cases.

Its development began when GPT-4.1 was the most advanced model available, the outlet says; it also reports that later testing suggested improved model capability does not necessarily eliminate the advantages of combining agents.

The sources therefore present both a notable test result and important boundaries around what it demonstrates: strong accuracy on the reported case sets, alongside a benchmark format that does not reproduce ordinary clinical practice.

The figures measure diagnostic correctness in those tests; they do not establish how the system would perform with patients or how clinicians would use it in care. Microsoft AI CEO Mustafa Suleyman framed the results as evidence that medical AI is arriving.

The Verge reports that he said the company wants to reach performance indistinguishable from human performance within a couple of years. TechCrunch carries a broader forecast from Suleyman: he said healthcare models could improve by an order of magnitude in capability within 24 to 36 months, and described the field as still being at an early stage.

TechCrunch also quotes Suleyman predicting that medical-record coding could be nearing a breakthrough in 2027. He said rapid progress can follow once a technology develops a “flywheel,” and described Microsoft’s commitment to accelerating healthcare AI as a possible way to distinguish it from other frontier AI companies.

These remarks set out the company’s ambitions, while the reported diagnostic results remain tied to specific test cases rather than demonstrated clinical outcomes.