A new Microsoft AI diagnosis study says its system got the right answer on difficult medical cases far more often than doctors working alone. The research, published on Thursday in the journal Science, tested MAI-DxO, Microsoft's "diagnostic orchestrator", on complex cases drawn from the New England Journal of Medicine. MAI-DxO was correct 83% of the time. Primary care doctors and general internists, working on their own with reference materials, were correct 36% of the time.
Mustafa Suleyman, chief executive of Microsoft AI, called the result a step toward what the company calls "medical superintelligence". The numbers are striking. The fine print matters just as much, and TechCrunch's Rebecca Bellan, who reported the study, laid much of it out.
How MAI-DxO works
MAI-DxO is not a single chatbot. It coordinates several AI agents that play the roles of a panel of doctors. Together they:
- question the "patient" to gather a history,
- order tests, each of which carries a cost in the simulation,
- argue over possible diagnoses before settling on one.
The setup rewards both accuracy and spending less on tests, which is closer to how hospitals actually work than a quiz-style benchmark.
Microsoft ran the orchestrator on top of models from OpenAI, Google, Anthropic, Meta and DeepSeek. Its best results came with OpenAI's GPT-5 and GPT-5.5.
The Microsoft AI diagnosis numbers in detail
According to TechCrunch's reading of the paper:
- GPT-5 on its own, with a short basic prompt, scored 78.9% on the complex cases at about $8,400 in simulated test costs per case. Adding MAI-DxO cut that cost by about $3,000 and raised accuracy by only about four points.
- GPT-5.5, released in April 2026, went from 72% to nearly 87% on complex cases with MAI-DxO, at a slightly lower cost.
- Emergency department cases showed little gain. GPT-5.5 scored about 85% with or without the orchestrator.
- Gemini 2.5 Flash, a smaller model, rose from 46.4% to nearly 69% with MAI-DxO.
That last pair shows the pattern. The orchestrator helps, sometimes a lot, but the gap between weaker and stronger models is bigger than the boost it adds. Microsoft's own researchers said gains shrink as models get close to the maximum possible score.
Three catches worth knowing
1. Specialists were not in the comparison
The 36% figure comes from primary care doctors and general internists working individually. The cases are the kind usually handled by specialists or teams of doctors, and neither group was tested. "Beats doctors" here means beats generalists working alone on very hard cases.
2. Most of the gain comes from the model
The data suggests the underlying model does most of the work. That matters for anyone deciding what to build or buy: a better model may deliver more than a clever wrapper.
3. It is a simulation
The cases are published teaching cases, not live patients. Real clinics have missing records, rushed consultations and patients who cannot describe symptoms clearly.
What happens next
Microsoft says the system is "ready for real world evaluations", in the words of Dominic King, vice president of health at Microsoft AI. Two projects are under way:
- Beth Israel Deaconess Medical Center, Boston: patients receive normal care while those who consent also interact with MAI-DxO in parallel. Researchers will compare its recommendations with those of the human doctors across emergency and urgent care.
- Mayo Clinic: Microsoft is working with Mayo to train a healthcare foundation model, with release planned for 2027.
MAI-DxO is not on sale. It remains a research and evaluation project. Google has published a similar real-world study of its AMIE diagnostic AI with Beth Israel in The Lancet, so the race between the big labs is now moving from benchmarks into hospitals.
What it means for you
For patients, nothing changes today. You cannot download MAI-DxO, and no regulator has cleared it for diagnosis. For health workers, the study is a signal that AI second opinions are coming, probably first as tools that sit beside a clinician rather than replacing one.
For health systems in Africa, where specialists are scarce and primary care carries most of the load, the comparison that matters is exactly the one Microsoft tested: generalists working alone on hard cases. If the Boston trial holds up, that is where tools like this could help most. It is also where the risks of a confident wrong answer are highest, so local validation, data protection and clear accountability will matter before any rollout.
Big companies are moving fast to put AI agents to work across whole organisations, as we saw with Uber's ChatGPT Work rollout. Medicine will move more slowly, and it should. The next number to watch is not 83%. It is how MAI-DxO performs on real patients in Boston.
