Microsoft's MAI-DxO beat doctors on complex diagnoses in peer-reviewed Science study

Microsoft's diagnostic AI outscored doctors on complex cases

A Microsoft-built diagnostic system correctly identified complex medical cases in a simulation more than twice as often as physicians, according to a study published in Science. Microsoft AI CEO Mustafa Suleyman described the result to TechCrunch as a step toward what the company calls "medical superintelligence," which Microsoft defines as the capacity to surpass the strongest contemporary human-led health systems in medical reasoning and coordination across the full arc of care.

The simulation used difficult cases from the New England Journal of Medicine. The system, MAI-DxO (DxO for short), reached the correct diagnosis 83% of the time. Primary care doctors and general internists, working individually with reference materials, were correct 36% of the time.

A panel of AI agents, not one model

MAI-DxO coordinates multiple AI agents that behave like a committee of clinicians: they interview the patient, order tests and debate a diagnosis before committing to one. According to TechCrunch, Microsoft ran the orchestration layer on top of models from several vendors, including OpenAI, Google, Anthropic, Meta and DeepSeek, but got its best results with OpenAI's GPT-5 and GPT-5.5.

Caveats behind the headline number

The comparison has limits that TechCrunch points out clearly: cases this complex are normally handled by specialists or physician teams, and neither group was included. The human baseline was generalists working alone.

The study's own data also suggests much of the performance comes from strong underlying models, with the orchestration adding gains on top. On its own with a short, basic prompt, GPT-5 scored 78.9% on complex cases at roughly $8,400 per case in estimated compute. DxO cut about $3,000 from that cost while adding only around four points of accuracy. On emergency department cases, DxO provided little improvement in accuracy or cost over GPT-5 alone.

Results split with GPT-5.5, released in April 2026. On complex cases, DxO lifted accuracy from 72% to nearly 87% at slightly lower cost, but on emergency cases both configurations scored about 85% at similar cost. The orchestration helps weaker models most: Gemini 2.5 Flash scored 46.4% alone and nearly 69% with DxO. Microsoft researchers said gains naturally shrink as models approach a theoretical ceiling, and a spokesperson framed DxO's contribution as optimizing both accuracy and cost.

Real-world evaluation comes next

Microsoft has announced a partnership with Beth Israel Deaconess Medical Center in Boston to evaluate DxO's clinical reasoning across hospital settings. Dominic King, vice president of Health at Microsoft AI, explained the design to TechCrunch: patients receive standard care, including triage, specialist time, investigations and discharge decisions, while consenting patients also engage with DxO in parallel. Microsoft and its partners will then compare the system's recommendations against the doctors' own. King said the study will span emergency and urgent care settings and predicted it will be the biggest of its kind.

Google, for its part, published a clinical study with Beth Israel in The Lancet this month evaluating its AMIE conversational diagnostic AI in real urgent care visits with physicians monitoring in real time. Microsoft now joins Google and OpenAI in claiming its AI can match or outperform doctors.

Suleyman, a DeepMind co-founder who joined Microsoft to lead its AI unit in 2024, also told TechCrunch that Microsoft is working with Mayo Clinic on a healthcare foundation model planned for release next year. He expects healthcare models to improve by an order of magnitude in capability within 24 to 36 months and predicts that in 2027 the field will see a "coding style breakthrough for medical records," with systems streaming accurate predictions of how a record will unfold.

Why it matters

This is peer-reviewed evidence, published in Science rather than on a company blog, that a multi-agent AI system can substantially outperform human clinicians at diagnostic reasoning, the core skill medicine is organized around. King frames the problem as the growing mismatch between demand for medical expertise and its supply; if DxO holds up in a Boston emergency department, the implications for triage and diagnosis are significant, and Microsoft already sells dictation and radiology tools to clinicians, giving it a deployment path. But the caveats matter as much as the milestone: the human benchmark excluded the specialists who would normally see these cases, the strongest results rely on OpenAI's models rather than Microsoft's own, and the emergency-department cases, the setting closest to real deployment, showed little benefit from the orchestration. The Beth Israel study is where those questions get answered.