Microsoft claims AI doctors outperform generalists on complex cases

A new study published this week in Science suggests that artificial intelligence can now diagnose difficult medical cases more accurately than human doctors. Microsoft’s latest system, known as MAI-DxO, correctly identified complex conditions in simulated scenarios more than twice as often as generalist physicians. Mustafa Suleyman, the CEO of Microsoft AI, described the results as a critical step toward what the company calls medical superintelligence.

The system works by coordinating multiple AI agents that function like a medical panel. These agents question the patient, order necessary tests, and debate potential diagnoses to reach a conclusion. In testing using complex cases from the New England Journal of Medicine, the AI was correct 83% of the time. By comparison, primary care doctors and general internists working alone with reference materials were correct only 36% of the time. It is worth noting that such complex cases are typically handled by specialists or teams in real life, groups that were not included in this specific comparison.

Dominic King, vice president of Health at Microsoft AI, told TechCrunch that the results address a significant gap in healthcare delivery. He noted that the current challenge is the growing mismatch between the demand for medical expertise and the available supply. With this release, Microsoft joins competitors like Google and OpenAI in asserting that their AI tools can match or exceed human performance in clinical reasoning.

The underlying data indicates that much of the accuracy gain comes from using more powerful AI models rather than the coordination system alone. Microsoft tested its system using models from OpenAI, Google, Anthropic, Meta, and DeepSeek. The best results were achieved using OpenAI’s GPT-5 and GPT-5.5 models. When GPT-5.5 was used, the DxO system raised accuracy on complex cases from 72% to nearly 87% while slightly reducing costs. However, on emergency department cases, the system provided little improvement over the standalone model, with both scoring around 85%.

Cost analysis showed mixed results. For complex cases, the orchestration system reduced the estimated cost per case by about $3,000 while improving accuracy by four points. For less capable models like Gemini 2.5 Flash, the system boosted accuracy from 46.4% to nearly 69%. Despite these gains, the gap between weaker and stronger models remained larger than the boost provided by the coordination layer.

Suleyman explained that Microsoft is collaborating with the Mayo Clinic to build a foundation model for healthcare, with a release planned for next year. He believes this infrastructure will enable major leaps in how medical records are handled. He predicted that by 2027, the company will see a coding-style breakthrough that allows for streaming highly accurate predictions about how a medical record might unfold, providing clinicians with valuable foresight.

Looking ahead, Microsoft is partnering with Beth Israel Deaconess Medical Center in Boston to evaluate the system in real-world settings. This study will compare DxO’s recommendations against those of human doctors across emergency and urgent care environments. Patients will continue to receive standard treatment, while consenting participants will also interact with the AI system. King stated that this will be the largest study of its kind and aims to provide compelling evidence for wider deployment.

The initiative coincides with a Microsoft manifesto on medical superintelligence, which is defined as surpassing the strongest human-led health systems in reasoning and coordination. Suleyman, who co-founded DeepMind before joining Microsoft in 2024, sees this focus as a key differentiator. He expects healthcare models to become an order of magnitude better in capability within the next 24 to 36 months, driven by continuous iteration and real-world usage data.