Microsoft AI Outperforms Doctors in Complex Diagnostics, Paving Way for Medical Superintelligence

Microsoft’s AI diagnostic system, known as MAI-DxO, has demonstrated the ability to correctly diagnose complex medical cases in simulations more than twice as often as generalist physicians. This finding, published Thursday in the journal *Science*, represents a significant step toward Microsoft CEO Mustafa Suleyman’s vision of achieving "medical superintelligence."

The system functions by coordinating multiple AI agents that simulate a panel of medical experts. These agents interact with patients to order tests and debate potential diagnoses. In a study utilizing complex cases from the *New England Journal of Medicine*, MAI-DxO achieved an 83% accuracy rate. In contrast, general practitioners and internists, who worked individually with reference materials, managed a 36% accuracy rate. Notably, the comparison did not include specialists or physician teams, as cases of this complexity are typically handled by them.

"There is something very substantial and exciting here," remarked Dominic King, Microsoft’s vice president of Health. He highlighted the growing mismatch between medical expertise supply and demand as a critical challenge in healthcare delivery.

By releasing these results, Microsoft joins competitors like Google and OpenAI in asserting that their artificial intelligence can match or exceed human diagnostic capabilities. Microsoft researchers attribute much of the improvement to advanced underlying models, with MAI-DxO’s orchestration system providing additional gains.

The system was tested using models from various providers, including OpenAI, Google, Anthropic, Meta, and DeepSeek, with the best outcomes achieved using OpenAI’s GPT-5 and GPT-5.5. Microsoft evaluated the system not only on accuracy but also on cost. For instance, running GPT-5 on its own resulted in a 78.9% score for complex cases at a cost of approximately $8,400. MAI-DxO reduced the cost by roughly $3,000 while slightly increasing accuracy. However, on emergency department cases, the improvement in accuracy or cost was minimal.

With the newer GPT-5.5 model, MAI-DxO demonstrated more significant improvements on complex cases, raising the accuracy score from 72% to nearly 87% at a comparable cost. Conversely, gains on emergency cases remained negligible. The system proved particularly useful for less capable models, such as Gemini 2.5 Flash, which saw its score jump from 46.4% to 69% with DxO.

Microsoft’s leadership believes this orchestration helps address the broader challenge of optimizing both accuracy and costs. Suleyman stated that the company is collaborating with the Mayo Clinic to train a healthcare-specific foundation model, with a release planned for next year. He views DxO as foundational infrastructure necessary for breakthroughs in healthcare.

Looking ahead to 2027, Suleyman predicts a "coding style breakthrough for medical records." He anticipates the ability to stream highly accurate predictions regarding how a medical record might evolve, offering clinicians valuable foresight. Suleyman, a co-founder of DeepMind who joined Microsoft in 2024, emphasized the "flywheel" effect, noting that rapid progress often follows when a model enters production for real-world iteration.

To test the technology in real-world settings, Microsoft will partner with Beth Israel Deaconess Medical Center in Boston. King described the upcoming study as the largest of its kind, evaluating clinical reasoning across emergency and urgent care environments to determine the system's potential for large-scale deployment.

This study coincides with Microsoft’s manifesto on "medical superintelligence," defined as the capacity to surpass contemporary human-led health systems in medical reasoning and coordination. Suleyman expressed commitment to accelerating the technology's introduction, viewing it as a differentiator for Microsoft in the competitive AI landscape.