Microsoft said on October 22, 2026, that its diagnostic-AI research had been published in Science, reporting substantially higher accuracy than individual generalist physicians on simulated medical cases. The findings concern a controlled diagnostic evaluation, not demonstrated improvements in patient care or evidence that AI can replace doctors. (microsoft.ai)
The company reported approximately 83% accuracy on complex cases, compared with 36% for physicians using reference resources. On emergency-department cases, the reported results were approximately 80% and 47%, respectively. Those comparisons depend on the cases selected, the resources available and whether the benchmark represents how clinicians actually work. (microsoft.ai)
A step-by-step diagnostic simulation
The evaluation used the Sequential Diagnosis Benchmark, or SDBench. According to the approved reporting, it comprised 304 clinicopathological conference cases from the New England Journal of Medicine and 325 emergency-department-to-admission encounters from MIMIC-IV-ED. These are benchmark totals, not verified counts of cases completed by each physician. (medicaleconomics.com)
Rather than receiving a complete case and immediately naming a diagnosis, participants began with a brief presentation. An AI gatekeeper supplied information as requested. The task therefore tested decisions about what to ask and which tests to request, alongside the eventual diagnosis. (medicaleconomics.com)
Microsoft’s Diagnostic Orchestrator, known as MAI-DxO or DxO, coordinates simulated physician roles that propose diagnoses, request evidence and challenge emerging conclusions. Its virtual panel is a software arrangement, not a team of independently practicing clinicians. The benchmark also tracks estimated testing costs, making the route to an answer part of the evaluation. (microsoft.ai)
What the physician comparison means
The human baseline was individual generalist physicians, rather than specialists or multidisciplinary teams. That distinction limits the conclusion: outperforming this comparator on selected scenarios does not establish superiority over the clinical teams that might investigate difficult cases in practice. The reported comparison does not measure the full capabilities of a healthcare system. (medicaleconomics.com)
Reference access also changes the apparent advantage. Microsoft reported approximately 22% physician accuracy on complex cases without external resources, versus 36% with them. Its near-fourfold comparison uses the lower baseline; against physicians with references, the reported advantage is approximately 2.3-fold. Neither comparison should be presented without identifying the conditions. (microsoft.ai)
The Verge’s account likewise distinguishes the retrospective case evaluation from clinical validation. The publication milestone is consequential, but publication alone does not turn a diagnostic simulation into evidence of safer treatment, better recovery or effective deployment in a hospital. (theverge.com)
How much does orchestration contribute?
A separate comparison is essential for developers: the orchestrated system versus its underlying language model. TechCrunch’s model-level figures, included in the approved source pack, report that GPT-5 reached 78.9% accuracy on complex cases with a short basic prompt. Adding DxO improved that result by about four percentage points. (techcrunch.com)
The reported benefit was larger with GPT-5.5 on complex cases, rising from 72% without DxO to nearly 87% with it. Yet emergency-department accuracy was approximately 85% in both configurations. Gemini 2.5 Flash reportedly rose from 46.4% to nearly 69% with orchestration. These comparisons remain secondary-source accounts pending inspection of the underlying study tables. (techcrunch.com)
Taken together, those figures suggest that neither a model-only explanation nor an orchestration-only explanation is sufficient. Strong underlying models account for substantial performance, while the added benefit of orchestration varies with the model and case set. For builders, the relevant question is the incremental gain over the same model without the orchestration layer—not simply the gap between the complete system and physicians. (techcrunch.com)
Accuracy and cost are separate results
Cost comparisons also require configuration details. Medical Economics reports that unconstrained DxO testing costs exceeded physician costs: $5,431 versus $4,725 on complex cases, and $4,979 versus $3,400 on emergency-department cases. These were estimated diagnostic testing costs, not verified AI-computing expenses or actual patient bills. (medicaleconomics.com)
Microsoft separately describes configurations optimized for lower costs. Consequently, a headline accuracy result and a headline cost saving cannot automatically be combined into one claim about the same operating setup. The meaningful comparison is a specified configuration’s accuracy alongside its testing expenditure. (microsoft.ai)
The next question is clinical benefit
The announcement builds on earlier research, rather than introducing the project for the first time. Microsoft’s linked account of that earlier work is dated June 30, 2025. Establishing exactly what changed between the earlier research and the reported Science publication still requires comparison of the final paper and supplementary material. (microsoft.ai)
Recent clinical research illustrates why patient-level evaluation matters. A separate trial published in Nature Medicine on June 26, 2026, enrolled 9,691 patients across 16 Kenyan primary-care facilities. It compared care with and without language-model assistance and found no statistically significant difference in its primary measure of treatment failure within 14 days. That result does not evaluate DxO, but shows why benchmark accuracy and clinical benefit are different questions. (nature.com)
Microsoft has announced a prospective evaluation with Beth Israel Deaconess Medical Center. That is a plan to collect clinical evidence, not a completed demonstration of effectiveness. For now, the supported conclusion remains narrower: the company reports strong diagnostic performance under benchmark conditions, with important dependencies on the physician comparator, underlying model and operating configuration. (microsoft.ai)
