Microsoft reports that its MAI-DxO diagnostic system correctly solved 83% of 304 complex medical cases, compared with 36% for generalist physicians allowed to consult reference materials. The results come from a text-based simulation and do not establish that the system delivers better care to patients.
In its October 22, 2026 announcement, Microsoft presents the research as progress toward “medical superintelligence.” The work, published in Science, supports a more specific finding: an AI system can outperform individual generalists at gathering information and diagnosing cases inside a controlled evaluation.
The study tests how a diagnostician chooses what to ask next. Whether that advantage survives contact with patients, clinical teams, and hospital workflows remains unresolved.
The 83% Result Depends on Who Took the Test
The headline comparison used 304 clinicopathological conference cases published in The New England Journal of Medicine. Microsoft reports 83% accuracy for MAI-DxO, versus 36% for generalists using textbooks and internet references, and 22% for physicians without those resources.
The reference-assisted result is the more informative comparison for readers assessing the headline. Access to additional information substantially changed physicians’ performance, though the AI still scored considerably higher.
A second evaluation used emergency-department-to-admission encounters drawn from MIMIC-IV-ED. Microsoft reports approximately 80% accuracy for MAI-DxO against 47% for physicians. This extends the evidence beyond unusually difficult published teaching cases, though it remains a retrospective simulation.
Neither comparison establishes superiority over specialist teams. As TechCrunch’s reporting notes, the complex NEJM cases would typically involve specialists or physician teams, whereas the study compared the AI with generalists working individually.
MAI-DxO outperformed the participating individual generalists under the study’s conditions. The result does not demonstrate that it outperforms the clinical organization that would ordinarily handle such cases.
SDBench Tests What the Diagnostician Chooses to Ask
SDBench, short for Sequential Diagnosis Benchmark, changes the task from answering a fully specified medical question to deciding how to investigate one.
According to the SDBench repository, the system under evaluation starts with a short case summary that withholds the diagnosis. It can ask questions, order tests, and eventually commit to a diagnosis. A separate AI “gatekeeper” supplies the requested information.
This structure tests whether the diagnostician gathers useful evidence before deciding, something a complete case write-up cannot show. A model handed all the decisive findings may recognize a diagnosis without demonstrating that it would have requested those findings itself.
Two implementation details qualify what the accuracy score measures.
First, the gatekeeper returns information from the underlying case when available, but can generate a synthetic response when the record lacks an answer. The evaluation therefore combines documented clinical evidence with simulated answers. The diagnostician’s decisions partly depend on how another model fills gaps in the source record.
Second, an AI judge scores the final diagnosis against the reference diagnosis on a five-point scale, counting scores of four or five as correct. Accuracy depends on both the diagnostic agent and the evaluation machinery interpreting its answer.
The public framework makes those mechanics inspectable, but access remains limited. The repository includes five synthetic development cases, not unrestricted access to the clinical datasets. NEJM cases require separate licensing, while the MIMIC dataset has separate access requirements. Publishing evaluation code helps researchers scrutinize the method; it does not establish independent reproduction of Microsoft’s results.
MAI-DxO’s Virtual Panel Is an Orchestration Layer
MAI-DxO coordinates several virtual-physician agents through repeated rounds of evidence gathering and debate. The agents maintain possible diagnoses, challenge leading explanations, and consider which tests to request.
These are software roles, not independent clinicians. Their value must be measured by whether the arrangement improves decisions over the underlying language model operating with a simpler setup.
TechCrunch reports that stronger underlying models account for much of the performance, with orchestration adding gains that vary by model and dataset. In its account, GPT-5 scored 78.9% on the complex cases with a basic prompt, while MAI-DxO raised accuracy by roughly four percentage points. The orchestration advantage was less apparent on emergency-department cases.
Attributing the entire 83% result to virtual debate would overstate its contribution. Model capability and the orchestration layer both contribute, and their relative importance changes across tasks.
Cost also depends on the configuration. Medical Economics’ study account reports that unrestricted MAI-DxO testing cost more than physicians’ test choices on average. Budget-constrained versions reduced estimated testing costs while sacrificing some accuracy.
Those estimates concern ordered investigations, not simply the computing bill for running the AI. They also do not capture every clinical consequence of testing, including discomfort, invasiveness, or time waiting for results. “More accurate and cheaper” should therefore describe a particular operating configuration, not the system in every setting.
The Hospital Study Must Answer a Different Question
Microsoft plans a prospective study with Beth Israel Deaconess Medical Center. According to TechCrunch’s description of the planned evaluation, consenting patients would interact with MAI-DxO alongside standard care, allowing researchers to compare its recommendations with clinicians’ decisions. That study has yet to demonstrate patient benefit.
The published benchmark evaluates text-only diagnosis. It does not establish treatment quality, patient safety, or outcomes, and it excludes parts of clinical practice such as interpreting images and dealing directly with incomplete or unreliable patient accounts.
Prospective evaluation lets researchers examine how the system’s recommendations behave in actual encounters, beyond recovering a known diagnosis from a simulated investigation. Whether clinicians can use those recommendations safely is a further question beyond diagnostic accuracy alone.
For the AI industry, the defensible advance is a more demanding, inspectable way to test diagnostic reasoning. Peer review strengthens the published evidence, but it does not turn a vendor-run benchmark into independently reproduced clinical effectiveness.
Microsoft’s “medical superintelligence” framing reaches beyond what these results establish. The planned hospital study must address the gap between a simulated diagnostic advantage and better patient care.
