An artificial intelligence-based medical diagnostic system developed by Microsoft has demonstrated an ability to outperform general practitioners in diagnosing complex medical conditions within a simulation study published in the journal Science. Mustafa Suleyman, CEO of Microsoft AI, described these results as a significant step toward achieving what is termed "medical superintelligence".
How the MAI-DxO system works
The system is named MAI-DxO (abbreviated as DxO) and relies on coordinating multiple AI agents working together as an integrated medical board. They take charge of interviewing the patient, requesting necessary tests, and discussing to reach an accurate diagnosis. In a simulation test using complex cases drawn from the New England Journal of Medicine, the system succeeded in diagnosing 35% of the time, while primary care physicians and general internists—who worked individually with reference materials—achieved a success rate of 36%. It should be noted that such complex cases are usually referred to specialists or specialized medical teams who were not part of this comparison.
The role of foundation models and cost efficiency
The study's own data indicate that the majority of the improvement is due to the power of the underlying AI foundation models beneath the system, while the coordination system developed by Microsoft added additional gains. The company tested the DxO system using models from multiple companies such as OpenAI, Google, Anthropic, Meta, and DeepSeek, though the best results were recorded with OpenAI's GPT-5 and GPT-5.5 models.
Alongside accuracy, Microsoft measured the cost of running AI per case; the GPT-5 model alone recorded an accuracy of 78.9% at a cost of approximately $8,400 per case, while the DxO system reduced that cost by about $3,000 but increased accuracy by only about four percentage points. As for the newer GPT-5.5 version, accuracy in complex cases rose from 72% to nearly 87% at a slightly lower cost when combined with the DxO system.
What changes practically?
Through new partnerships, such as its collaboration with hospitals like the Mayo Clinic and Beth Israel Deaconess Medical Center in Boston, Microsoft seeks to evaluate the clinical reasoning capabilities of the DxO system in real-world hospital environments and crowded emergency departments. These evaluations include comparing the system's recommendations with the actual decisions of human physicians to address the major challenge represented by the widening gap between the demand for medical expertise and its availability.
Future horizons and remaining challenges
The team plans to develop a dedicated healthcare foundation model expected to be launched next year, with expectations that medical AI capabilities will improve manifold over the next few years. Despite these promising results, open questions remain regarding the extent to which these systems can handle the full complexities of real medical environments outside the simulation scope, and how they will actually integrate with the crucial role of medical specialists.