Microsoft's MAI-DxO diagnostic system now has the credential its 2025 preprint lacked: publication in the peer-reviewed journal Science on October 22, 2026. According to Gizmodo, the system reached 83.2% accuracy on complex New England Journal of Medicine cases. Generalist physicians with reference materials scored 36.1%, and without them they fell to roughly 22%.
On emergency department cases, the system scored 80.3% against 46.8% for generalist physicians, according to Unite.AI. Mustafa Suleyman, CEO of Microsoft AI, described the result as a step toward "medical superintelligence," as TechCrunch reported.
The publication landed on a rough day for AI stocks. BNN Bloomberg reported that AI shares fell worldwide on October 22, with borrowing-cost worries about data centers and chips weighing on the group while oil prices also climbed. The study has nothing to do with that sell-off, but both stories will reach many of the same readers.
How MAI-DxO works and which models power it
MAI-DxO does not produce a diagnosis in one shot. It works sequentially, asking questions, ordering tests and then reaching a conclusion, and according to the reporting it ordered fewer tests than physicians did. That design mirrors how clinicians gather information, and it sits at the center of Microsoft's argument that the system is more than a large model answering a medical quiz.
The 2025 version pooled several off-the-shelf models, including Claude, DeepSeek, Gemini, GPT, Grok and Llama, under a coordinating orchestrator, as TIME reported in its 2025 coverage. The Science version relies on OpenAI's models. Its strongest results used GPT-5, and further gains appeared with GPT-5.5. OpenAI's newer GPT-6 models are already in preview on Google Cloud, a reminder of how quickly the model layer underneath these results is moving. Our earlier report on that rollout has the details.
That detail changes how the headline should be read. TechCrunch's analysis suggests most of the improvement comes from the underlying AI models rather than from Microsoft's orchestration layer. If that holds, the study measures frontier model progress as much as it validates Microsoft's architecture. How much credit the orchestration deserves is still being debated.
Accuracy and cost against physicians
Accuracy is only part of what the paper reports. Cost per case also appears in the results, and it will shape how hospital finance teams judge the tool.
| Setting | MAI-DxO | Generalist physicians |
|---|---|---|
| Complex NEJM cases, unconstrained | 83.2% at about $5,431 per case | 36.1% with reference materials at about $4,725; roughly 22% closed-book at about $3,049 |
| Complex NEJM cases, cost-constrained | 79.6% at $1,303 per case | Not applicable |
| Emergency department cases | 80.3% at $4,979 per case | 46.8% at $3,400 per case |
Two numbers stand out. The best single physician on the NEJM set reached 51.8%, so the system beat the strongest individual doctor by more than 30 points. The cost-constrained configuration kept accuracy near 80% at $1,303 per case, which suggests the tradeoff between accuracy and cost can be tuned. But the unconstrained system cost more per case than physicians with reference materials, so the economic case is not automatic.
What the study does not show
The biggest caveat is the comparison group. The study pitted the AI against generalist physicians, not the specialists who usually handle the hardest and rarest cases, as CryptoBriefing noted. A generalist is a fair baseline for many emergency workflows, but it is not the ceiling of human diagnostic skill.
Second, these were case-based evaluations, not trials with real patients in real clinics. Critics note that benchmark wins do not automatically translate into better care. Zetik's report says deskilling and "mis-skilling" are real risks when clinicians lean on AI, so a tool that scores well on paper could still change how doctors reason.
Third, the numbers are not yet settled. The 2025 preprint reported 85.5% accuracy on NEJM cases, and the Science version comes in slightly lower at about 83%. The physician baselines also differ, with about 20% for physicians in the preprint against 36.1% in the Science paper, so the two results do not form a clean before-and-after comparison.
Sample descriptions conflict as well. One account cites about 304 NEJM teaching cases and 325 emergency admissions, according to CellCog's analysis. Another describes a physician comparison on 56 held-out cases. The coverage does not reconcile those figures, so check the paper's methods section before citing a sample size.
When hospitals could test MAI-DxO
Microsoft plans real-world evaluation with the Mayo Clinic and Beth Israel Deaconess Medical Center in Boston. The plan calls for running the AI alongside clinicians in urgent and emergency settings, according to CryptoBriefing and Whalesbook. Microsoft frames the tool as support for clinicians, not a replacement for them.
The timeline is vague. Sources describe the trials as coming in the "coming months," with no confirmed start date. The sources reviewed also do not say whether Microsoft will seek FDA clearance or another regulatory review. That gap matters. A tool that helps guide decisions in emergency departments will draw hard questions about validation, liability and oversight, and the answers will shape whether hospitals adopt it.
For US patients, nothing changes today. The tool is not in routine care, and the sources reviewed include no real-world results. For clinicians and hospital executives, the study raises a sharper question: whether AI diagnostic support belongs in urgent and emergency workflows at all, and what oversight it would need if it does.
Why investors are watching Microsoft's medical push
Microsoft's medical work depends on OpenAI's models, because the strongest Science results used GPT-5 and GPT-5.5. Investors following healthcare AI will watch whether that partnership produces clinical products, not just published papers. The timing also sits against a broader debate about data center costs. President Trump has argued that data centers should generate their own power, as our earlier report explains, and that argument feeds directly into the borrowing-cost worries that hit AI stocks this week.
What clinicians and patients can do now
No action is needed for individual care today. A few steps still make sense.
- Patients: ask your clinician whether your hospital uses AI diagnostic support, and how a human reviews its suggestions before they affect your treatment.
- Patients: if you want an AI-assisted second opinion, ask how the tool was validated and whether it was tested on patients like you.
- Clinicians and administrators: before any pilot, document who signs off on AI-suggested diagnoses, how disagreements are logged and how errors are reported.
- Anyone citing the study: check the Science paper's methods section for sample sizes, because the coverage disagrees.
What to watch next
Three things will show whether this result holds beyond the lab. The first is the start date and design of the Mayo Clinic and Beth Israel Deaconess studies, especially whether they measure patient outcomes instead of accuracy scores alone. The second is any regulatory filing, since the sources do not say one is planned. The third is independent teams testing the system with newer models, which would show whether the gains survive outside Microsoft's research group.
The Science publication is a credibility milestone for a result first announced as a preprint. The harder milestone is a study in which real patients and real clinicians are part of the test.
