Microsoft MAI-DxO turns diagnosis into a debate among a panel of virtual physicians. On October 22, 2026, Microsoft AI published a paper on its official blog introducing two things at once: the Sequential Diagnosis Benchmark (SDBench) and MAI-DxO (Microsoft AI Diagnostic Orchestrator), the leading AI system on that benchmark. On 304 complex cases selected from the New England Journal of Medicine, MAI-DxO reached 83% diagnostic accuracy — about 3.7 times the physicians' 22%.
The numbers first: three comparisons tell the same story
| Scenario | MAI-DxO | Physicians |
|---|---|---|
| NEJM complex cases | 83% | 22% (no external resources) / 36% (with references) |
| Typical ED cases (MIMIC-IV-ED) | 80% | 47% |
| Cost-optimized config | 81%, at less than half the test cost of comparison systems | — |
Microsoft stresses that across both datasets, no other human or AI approach was both more accurate and less costly.
The key is not the model, but the 'consultation' workflow
MAI-DxO doesn't rely on one model for a single-shot answer: a panel of 'virtual physicians' each proposes diagnoses, picks questions and tests, then challenges the emerging view and updates the assessment over several debate rounds; every test is weighed for information value against cost before being ordered.
SDBench's design is equally key: it makes cases interactive — the AI or physician starts from limited information and decides what to ask and check next, while the benchmark scores accuracy, evidence requested and estimated costs. It tests the 'diagnostic path', not just the final answer.
What was open-sourced: part of the benchmark code is out
Microsoft is open-sourcing SDBench for select NEJM cases (benchmark, Gatekeeper, Judge and cost-estimation code) so researchers can reproduce or challenge the findings; re-evaluated with newer models (including GPT-5.5), the gains persisted, suggesting the orchestration is not tied to one model generation.
Boundaries: paper results are not clinical readiness
The NEJM cases are selected complex cases, not everyday outpatient visits; the physicians' 22% was measured without external resources (36% with references). Cost estimates are model-based; real healthcare cost structures are far more complex. The work is framed as a paper — the next step is more rigorous evaluation, not direct clinical deployment.
What researchers can do now
Reproduce MAI-DxO's results, run your own model through it, or probe the 'accuracy + cost' dual evaluation for holes. For medical AI teams, the more transferable lesson may be methodological: as model gaps narrow, multi-model orchestration is becoming a new source of performance.
