Microsoft AI Claims 83% Accuracy In Complex Case Diagnosis, Outpacing Physicians

Microsoft AI is pushing a bold narrative this week: its new diagnostic system, MAI-DxO, is designed to outperform human physicians when handling complex medical cases. In a simulation detailed in a study published Thursday in Science, the system achieved an 83% accuracy rate on difficult diagnoses. For comparison, primary care doctors and general internists working individually with reference materials were correct only 36% of the time. Mustafa Suleyman, CEO of Microsoft AI, framed the results as a critical step toward building what he calls medical superintelligence.

The timeline of this development begins with the architecture behind MAI-DxO, known as DxO. The system does not rely on a single algorithm. Instead, it coordinates a panel of AI agents that simulate the workflow of a medical team. These agents question the patient, order tests, and debate the diagnosis before reaching a conclusion. The study utilized complex cases drawn from the New England Journal of Medicine for the simulation. It is important to note that cases of this complexity are typically handled by specialists or physician teams, who were not included in this specific comparison group.

Microsoft joins Google and OpenAI in the race to claim AI superiority in healthcare, but the company's approach involves specific technical trade-offs regarding cost and model performance. While Microsoft tested DxO with models from several companies, including Anthropic, Meta, and DeepSeek, the strongest results appeared when utilizing OpenAI's GPT-5 and GPT-5.5. The study indicates that much of the diagnostic improvement stems from the underlying power of these large language models, with Microsoft's orchestration system providing additional gains on top of that foundation.

The data also highlights where the system adds value and where it does not. Microsoft measured the estimated cost of running the AI for each case. On its own, GPT-5 scored 78.9% on complex cases at a cost of approximately $8,400 per case. Using DxO reduced that cost by about $3,000 while raising accuracy by only about four points. However, on emergency department cases, there was little improvement in either accuracy or cost when using DxO compared to running GPT-5 alone.

Results varied further when Microsoft tested its own GPT-5.5 model, released in April 2026. For complex cases, DxO increased the model's accuracy from 72% to nearly 87% at a slightly cheaper cost. Yet, the gains did not materialize for emergency department cases, where both the standard model and DxO scored about 85% at a similar cost. Microsoft researchers acknowledged that accuracy gains naturally shrink as models approach the theoretical maximum, noting that DxO helps with the broader challenge of optimizing both accuracy and cost.

Looking ahead, Microsoft is expanding its efforts into real-world clinical settings. The company has partnered with Beth Israel Deaconess Medical Center in Boston to evaluate the clinical reasoning capabilities of DxO across different hospital settings. The study will involve patients receiving standard treatment while also engaging with DxO, allowing partners to compare the AI's recommendations against those of human doctors. Dominic King, vice president of Health at Microsoft AI, stated that the company believes this will be the biggest study of its kind across emergency and urgent care environments.

Suleyman is focused on the long-term horizon, specifically the year 2027. He stated that he expects healthcare models to get an order of magnitude better in capability within the next 24 to 36 months. He believes that in 2027, the industry will be on the verge of a breakthrough for medical records, allowing systems to stream highly accurate predictions about how a medical record might unfold. Through a partnership with the Mayo Clinic, Microsoft is working to train a foundation model for healthcare with a release planned for next year.