83%: The Accuracy Gap in Microsoft's New AI Diagnostic Panel

Microsoft’s AI diagnostic system, MAI-DxO, has demonstrated an ability to correctly identify complex medical conditions in simulation 83% of the time, a result that more than doubles the performance of individual generalist physicians. Published in the journal Science on Thursday, the study positions the technology as a critical step toward what CEO Mustafa Suleyman describes as “medical superintelligence.” The system operates by coordinating multiple AI agents that function like a medical panel, engaging in patient questioning, ordering tests, and debating potential diagnoses.

The comparison baseline was specific: primary care doctors and general internists working individually with reference materials achieved a 36% accuracy rate on the same set of complex cases derived from the New England Journal of Medicine. However, the study notes that such difficult cases are typically managed by specialists or teams in practice, who were not included in this direct comparison. Dominic King, Vice President of Health at Microsoft AI, characterized the findings as substantial, noting that the primary challenge in healthcare is the widening gap between medical demand and the supply of expertise.

Analysis of the study’s data indicates that much of the performance improvement stems from the power of the underlying foundation models rather than the orchestration layer alone. Microsoft tested the system with models from various providers, including OpenAI, Google, Anthropic, Meta, and DeepSeek. The highest accuracy was achieved using OpenAI’s GPT-5 and GPT-5.5 models. When GPT-5 was used without the DxO orchestration, it scored 78.9% at an estimated cost of $8,400 per case. The addition of DxO reduced that cost by approximately $3,000 but increased accuracy by only about four points.

The benefits of the orchestration system were more pronounced when paired with newer or less capable models. With GPT-5.5, released in April 2026, DxO raised accuracy from 72% to nearly 87% at a slightly lower cost. Conversely, the system showed minimal improvement for emergency department cases, where both configurations scored around 85% at similar costs. For weaker models like Gemini 2.5 Flash, DxO lifted accuracy from 46.4% to nearly 69%, yet the gap between strong and weak models remained larger than the boost provided by the orchestration. Researchers noted that accuracy gains naturally diminish as models approach theoretical maximums.

Suleyman, who joined Microsoft in 2024 after co-founding DeepMind, 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 anticipates that within 24 to 36 months, healthcare models will become an order of magnitude better in capability. To validate these capabilities in real-world settings, Microsoft has partnered with Beth Israel Deaconess Medical Center in Boston. This upcoming study will compare DxO’s recommendations against human physicians during urgent care and emergency visits, aiming to provide compelling evidence for large-scale deployment.

This initiative coincides with a broader manifesto from Microsoft defining “medical superintelligence” as the capacity to surpass human-led health systems in reasoning and coordination. While the company already sells dictation and radiology tools, DxO represents an expansion into clinical decision support. Suleyman expressed a belief that a “coding style breakthrough” for medical records is imminent in 2027, allowing for highly accurate predictions of patient outcomes. He emphasized that the company is committed to becoming the leading provider of state-of-the-art medical AI, viewing this focus as a key differentiator among frontier labs.