MAI-DxO is the diagnostic system Microsoft published in the journal Science on October 22, 2026. Instead of having one model answer directly, it organizes a panel of virtual physicians that debate and verify step by step. According to Microsoft's official introduction the same day, the system reached 83% diagnostic accuracy on complex cases from The New England Journal of Medicine, against 36% for generalist physicians with access to reference materials and just 22% without.
First, change the exam to look like real diagnosis
Microsoft argues that licensing-exam-style tests overrate AI: questions present all information at once and ask for a single answer, while real diagnosis starts from limited information and requires deciding what to ask and what to test, with every additional test carrying a cost. The team therefore built SDBench: 304 complex NEJM cases plus a set of more typical emergency-department cases from the MIMIC-IV-ED database, all converted into interactive encounters. The diagnostician starts with a short presentation and must request each piece of evidence, while the benchmark records accuracy, the evidence requested and estimated test costs — judging the path, not only the answer.
A panel of virtual doctors investigates better than one model
MAI-DxO has multiple agents each propose diagnoses, select questions and tests, and challenge one another over several rounds of debate, updating the assessment as evidence arrives and weighing each test's likely value against its cost. The orchestration is not tied to one model: with the newer GPT-5.5 underneath and no change to the method, results improved further. On emergency-department cases it scored 80% against physicians' 47%, and a cost-focused configuration kept 81% accuracy on NEJM cases at less than half the testing cost of comparison systems. Microsoft has open-sourced the SDBench tooling for a selection of NEJM cases, including the benchmark, gatekeeper, judge and cost-estimation code, so other teams can reproduce or challenge the findings.
The limits deserve equal clarity: this is a controlled text benchmark, the comparison group is generalist physicians working alone, and rare complex cases in the real world are normally referred to specialist teams; the system has not been validated in real clinical workflows. Reading it as "AI can already replace doctors" is a misread. The better reading is that diagnostic evaluation is moving, as with Google AMIE's real-clinic study, from answering questions correctly to asking the right questions at the right cost — and that multi-agent orchestration shows a real edge on this kind of process-heavy task.
