Microsoft AI System Outperforms Doctors on Complex Medical Cases

A study published Thursday in the journal Science claims that Microsoft’s new AI diagnostic system correctly identified complex medical cases in a simulation at more than twice the rate of human physicians.

The system, known as MAI-DxO, utilizes multiple AI agents to simulate a panel of medical experts. These agents question a patient, order necessary tests, and debate potential diagnoses to reach a conclusion. In the simulation, which featured complex scenarios from the New England Journal of Medicine, the system achieved an 83% accuracy rate. By comparison, primary care doctors and general internists, working individually with reference materials, were correct 36% of the time.

Mustafa Suleyman, the CEO of Microsoft AI, characterized the result as a step toward “medical superintelligence,” which the company defines as the capacity to surpass the strongest contemporary human-led health systems in medical reasoning and coordination across the full arc of care. Dominic King, vice president of Health at Microsoft AI, described the results as “very substantial and exciting,” noting that the biggest challenge in healthcare delivery is the growing mismatch between the supply and demand of medical expertise.

However, the study includes significant caveats. The cases used in the comparison are typically managed by specialists or teams of physicians, who were not part of this specific test. The research also indicates that the vast majority of the improvement stems from powerful underlying AI models, with Microsoft’s orchestration system providing additional gains on top of that baseline.

Microsoft tested its DxO system using models from various providers, including OpenAI, Google, Anthropic, Meta, and DeepSeek. The highest accuracy was recorded when using OpenAI’s GPT-5 and GPT-5.5. The study also analyzed the estimated cost of running the AI for each case. For example, running GPT-5 alone on complex cases scored 78.9% accuracy at approximately $8,400 per case. Using DxO reduced that cost by about $3,000 while raising accuracy by roughly four points.

Results varied depending on the specific scenario. On emergency department cases, there was little improvement in either accuracy or cost when adding DxO to GPT-5. With GPT-5.5, a model released in April 2026, DxO increased accuracy on complex cases from 72% to nearly 87% at a slightly lower cost. However, on emergency department cases, both the standard model and the DxO-assisted version scored about 85% at a similar cost. DxO provided the most assistance to less capable models, such as Gemini 2.5 Flash, which scored 46.4% on its own but reached nearly 69% with DxO assistance.

Suleyman stated that Microsoft is collaborating with the Mayo Clinic to train a foundation model for healthcare, with a release scheduled for next year. He believes this model will enable the company to “make really enormous leaps here,” and suggested that DxO is part of the foundational infrastructure Microsoft is building to achieve breakthroughs in healthcare. “I really believe that in 2027, we will be on the verge of a coding style breakthrough for medical records,” Suleyman said. He added that he expects healthcare models to improve “an order of magnitude better in capability within the next 24 to 36 months.”

Microsoft is initiating a partnership with Beth Israel Deaconess Medical Center in Boston to evaluate the clinical reasoning capabilities of DxO in real-world hospital settings. During the evaluation, patients will receive standard treatment, while consenting participants will also engage with DxO. Microsoft and its partners will then compare the AI’s recommendations against the decisions made by human doctors. Suleyman and King have both expressed confidence that this technology is ready for real-world evaluation and can assist clinicians in making accurate predictions about how a medical record might unfold.