Microsoft AI Outperforms Physicians in Complex Cases

Microsoft's AI diagnostic system correctly solved complex medical cases in a simulation at more than twice the rate of human doctors, according to a study released Thursday. The results show a significant leap in machine capability for difficult diagnoses.

In tests using complex cases from the New England Journal of Medicine, the system, called MAI-DxO, achieved an accuracy of 83 percent. By comparison, primary care doctors and general internists working individually with reference materials were correct 36 percent of the time. It is important to note that cases of this complexity are typically handled by specialists or physician teams, who were not part of the comparison group.

The system works by coordinating a panel of AI agents that act like a group of doctors. This virtual panel questions a patient, orders tests, and debates a diagnosis to reach a conclusion. Mustafa Suleyman, the CEO of Microsoft AI, described the result as a step toward what he calls medical superintelligence.

The study suggests that much of the accuracy comes from powerful underlying AI models. Microsoft tested its orchestration system with models from several companies, including OpenAI, Google, Anthropic, Meta, and DeepSeek. The top results were achieved using OpenAI's GPT-5 and GPT-5.5 models.

Microsoft also measured the system on cost as well as accuracy. On its own with a simple prompt, GPT-5 scored 78.9 percent on complex cases at an estimated cost of about $8,400 per case. The DxO system cut that cost by about $3,000 while raising accuracy by about four points. The system's performance on emergency department cases showed little change in either accuracy or cost when compared to the GPT-5 model operating on its own.

With GPT-5.5, a newer model released in April 2026, the results were mixed. On complex cases, DxO increased the accuracy score from 72 percent to nearly 87 percent at a slightly cheaper cost. However, there were no significant gains on emergency department cases, where both versions scored around 85 percent at a similar cost.

The system provides the most help on complex cases and when working with less capable AI models. For example, the Gemini 2.5 Flash model scored 46.4 percent on its own but reached nearly 69 percent when assisted by DxO. Even so, the difference in performance between weaker and stronger models was larger than the improvement provided by the DxO system.

The company is now moving to test its system in real-world settings. Through a partnership with Beth Israel Deaconess Medical Center in Boston, Microsoft will evaluate the clinical reasoning capabilities of DxO across different hospital settings. Patients will receive standard treatment from human doctors. In parallel, consenting patients will also engage with the DxO system, and Microsoft will compare its recommendations against those of human clinicians.

This effort is part of a broader goal to build what the company defines as medical superintelligence, described as the capacity to surpass the strongest current human-led health systems in medical reasoning across the full arc of care. Suleyman, who previously co-founded DeepMind, said that healthcare models will improve an order of magnitude in capability within the next 24 to 36 months. He believes that in 2027, the company will be on the verge of a breakthrough for medical records that can stream highly accurate predictions about how a record might unfold.