The topline. San Juan, October 23, 2026. Every clinician knows the case that turns on the question nobody asked. A study published in Science this week tested whether an AI that works cases the way clinicians do, sequentially, asking questions and ordering tests one step at a time, could be the colleague who asks it. Microsoft’s Diagnostic Orchestrator reached the correct diagnosis in 83.2 percent of NEJM clinicopathological conference cases, against 36.1 percent for practicing generalist physicians with reference tools. Read as its authors intend, the result is not a verdict on physicians. It is a measurement of the headroom above a solo workup, and of the errors a tireless second reader can catch before they set.

Key facts

The system is the Diagnostic Orchestrator, or DxO, built by Microsoft researchers on a commercial large language model, GPT-5 in the main results. Instead of producing one answer, it simulates a panel of physicians with different roles who track a differential, challenge the leading hypothesis, and weigh the cost of each test before ordering it.

The researchers built a benchmark called SDBench from 304 New England Journal of Medicine clinicopathological conference cases and 325 emergency department to admission encounters from the MIMIC-IV-ED database. A separate AI gatekeeper released case information only when specifically requested, and each test ordered carried an estimated cost from hospital chargemaster pricing. Earlier studies of AI diagnosis had a familiar weakness: they handed the model a complete, tidy case write-up in a single prompt and graded the final answer. That is not how diagnosis works. Patients arrive with a fragment, and the clinician builds the picture one question and one test at a time.

On the emergency department cases, DxO scored 80.3 percent against 46.8 percent for physicians working alone. The highest scoring individual physician reached 51.8 percent on the NEJM set. These are cases selected to be hard; the distance between a solo workup and the second reader’s result is exactly the territory where diagnostic errors live.

Clinical relevance

Working styles differed in a way that teaches. Physicians asked an average of 5.6 questions before ordering their first test; DxO asked 25.5, more extensive structured history-taking. DxO’s misses clustered near the answer, a closely related diagnosis at the wrong level of specificity, while physician misses landed further from the correct diagnosis, the pattern a second reader exists to interrupt.

Cost is where the nuance sits. Without budget limits, DxO spent more per case. But budget-constrained versions cut costs while giving up little accuracy: 80.9 percent accuracy at $2,241 per case on the NEJM set, still above the physician average.

The caveats deserve attention too. The comparison was against individual generalists working alone, not specialists or multidisciplinary teams, which is precisely the setup a second reader is meant to change. The benchmark was text only, excluded the messiness of real encounters, and measured only diagnosis, not treatment or outcomes. All authors are Microsoft employees holding company equity.

83.2%

That is DxO’s accuracy on the NEJM cases, the hardest public diagnostic benchmark in medicine, against 36.1 percent for generalists using search and reference tools and 22.3 percent without them. The number matters less than the setup that produced it: this was the first major benchmark that forced the model to earn its answer the way a clinician earns one. The authors’ next step is designing workflows where clinicians and AI work together. If the panel in the machine asks better questions and knows its budget, its place is beside the physician, not across from them: a second panel to argue with, so the physician’s own judgment lands more often.

What we’re reading

  • Science: the DxO paper itself, with the SDBench methodology in full.
  • Medical Economics: the clearest clinical read of what the results mean for practice.
  • NEJM: the clinicopathological conferences behind the benchmark’s case set.
  • Microsoft Research: the project’s framing of orchestrated diagnostic agents.