Microsoft says its MAI-DxO diagnostic AI correctly solved 83% of complex New England Journal of Medicine (NEJM) cases on a benchmark, against 22% for physicians working without references. In a post titled "Building Medical Superintelligence", dated Oct. 22, 2026, Microsoft's MAI Health Team says the peer-reviewed paper is published today in Science. It is a research result, not a product: the post names no clinical deployment and no availability.

What Microsoft reports

All figures below are as Microsoft states them in its post.

  • Complex NEJM cases: MAI-DxO 83%, physicians 22% without references and 36% with them. Microsoft describes that as about 3.7 and 2.3 times the physicians' results.
  • Emergency-department cases (from the MIMIC-IV-ED dataset): MAI-DxO 80% against 47% for physicians, which Microsoft calls "more than 1.7 times."
  • Cost: Microsoft says one cost-optimised configuration reached 81% on the NEJM cases "at less than half the cost of comparative systems." The post also claims the system costs half to a third as much, with fewer tests.

The tests run on SDBench, the Sequential Diagnosis Benchmark, which Microsoft publishes online. The paper is at Science, linked from Microsoft's post, and a companion paper, "The Building Blocks of Medical Superintelligence," is in BMJ Digital Health and AI this week, according to Microsoft.

What it builds on

This is the peer-reviewed version of work Microsoft first put out as a June 2025 preprint, "The Path to Medical Superintelligence." The post presents the Science publication and the emergency-department results.

"Medical superintelligence" is Microsoft's own term, and we use it only in quotation. Microsoft also says it is merging its health teams, including Microsoft Research, Dragon Copilot and PowerScribe, into Microsoft AI.

What the numbers do not show

  • The benchmark uses case records, not live patients, so the percentages are not a measure of how the system would do in a clinic.
  • Microsoft's post describes the physicians' conditions only as with or without reference resources such as textbooks and the internet.

For other recent AI results reported on benchmarks, see OpenAI's GPT-6.1 Sol COBOL migration score and Hugging Face's Carbon-X DNA model.