Microsoft has a diagnostic system that, in a simulation of complex cases, was right more than twice as often as primary care doctors. The number in the headlines is 83 percent against 36. The more interesting number is in the company's own analysis: much of the lift does not come from the system. It comes from the model underneath it.
The system is called MAI-DxO. It orchestrates several AI agents that behave like a panel of doctors — questioning the patient, ordering tests, arguing over differentials, converging on a diagnosis. It was tested on complex cases from the New England Journal of Medicine, which is to say the kind of case real hospitals hand to specialists or whole teams. The study was published in Science.
What the numbers show
The panel scored 83 percent. Primary care doctors and general internists, working alone with reference materials, scored 36 percent. That comparison comes with a caveat the study itself makes: specialists and physician teams were not in the control group.
Then it gets interesting. Microsoft tested DxO with models from several vendors — among them OpenAI, Google, Anthropic and Meta — and got its best results with GPT-5 and GPT-5.5. GPT-5 alone, given a short baseline prompt, scored 78.9 percent on the complex cases. DxO cut the estimated cost per case by about $3,000 and raised accuracy by roughly four points. On emergency department cases, there was little difference in either accuracy or cost between the model alone and the model with DxO.
With GPT-5.5 the picture split: on complex cases, accuracy rose from 72 percent to nearly 87 percent at slightly lower cost; in the emergency department, again no difference, both around 85 percent.
The effect is largest on weaker models. Gemini 2.5 Flash scored 46.4 percent alone and nearly 69 percent with DxO. But the gap between weak and strong models was bigger than the gain the orchestration delivered. Microsoft's researchers say plainly that accuracy gains shrink as a model approaches the theoretical maximum.
A cost figure worth sitting with
The study also prices running the AI per case. GPT-5 alone on a complex case: about $8,400. That is not a product price, it is a bill for model calls in a multi-step procedure — and it shows how far this architecture sits from a clinic. DxO reduces that figure, but it remains an order of magnitude that only makes sense in a study where every answer counts.
The company's actual argument
Mustafa Suleyman, chief executive of Microsoft AI, calls the result a step toward "medical superintelligence", defined as the capacity to surpass the strongest contemporary human-led health systems in medical reasoning and coordination across the full arc of care. That is deliberately not "beats a doctor on an exam question". It is "beats a system".
Which is why the next step matters more than the headline. With the Mayo Clinic, Microsoft plans to train a healthcare foundation model, targeted for next year. And with Beth Israel Deaconess Medical Center in Boston it is starting a real-world study: patients receive standard care, and those who consent also engage with DxO in parallel, with the recommendations compared. Dominic King, vice president of health at Microsoft AI, calls it the largest study of its kind — emergency and urgent care, not just laboratory conditions.
Google published a Lancet study this month with Beth Israel on its conversational diagnostic AI, AMIE, with physicians monitoring interactions in real time. The bar for "medical superintelligence" has therefore moved from benchmarks to controlled clinical trials.
What is missing before you believe it
This is a simulation experiment on published cases, not on patients. A case that appears in a journal has already been written down, already triaged, and already contains the information that leads to the answer — an advantage a real first encounter does not offer. There are no figures for wrong-but-dangerous answers, for over-investigation, or for how the system handles incomplete or incorrect information. And comparing against generalists working alone measures the gap the system is meant to fill, not its ability to fill it.
Suleyman's forecast is bold and datable: in 2027, he says, the field will be on the verge of a "coding-style breakthrough" for medical records, streaming predictions about how a case might unfold. Whether that happens will not be decided by 83 percent. It will be decided by the study that has only just begun.
