Microsoft MAI-DxO Beats Generalist Doctors 83% to 36% in Complex Diagnosis Test
Microsoft's MAI-DxO diagnosed 83% of complex cases versus physicians' 36%, but hospital trials must show whether simulated gains translate to real clinical care.
Summary
On October 22, 2026, Microsoft said MAI-DxO, a multiagent system that questions patients, orders tests and debates diagnoses, correctly solved 83% of complex New England Journal of Medicine cases in simulation. Generalist primary care doctors and internists working individually with references scored 36%, although specialists and physician teams that normally handle such cases were excluded. Microsoft AI CEO Mustafa Suleyman called it progress toward “medical superintelligence” to relieve expertise shortages, joining similar Google and OpenAI claims.
Microsoft tested models from OpenAI, Google, Anthropic, Meta and DeepSeek, with GPT-5 and GPT-5.5 performing best. GPT-5 alone scored 78.9% at about $8,400 per complex case; DxO saved about $3,000 and added only about four points, with little emergency department improvement. GPT-5.5, released in April 2026, rose from 72% to nearly 87% with DxO at slightly lower cost, but both versions scored about 85% on emergency cases at similar cost. Gemini 2.5 Flash climbed from 46.4% to nearly 69%, showing orchestration helps weaker models, although underlying model strength mattered more and gains shrink near maximum accuracy.
Microsoft and Mayo Clinic are developing a healthcare foundation model due in 2027; Suleyman predicts a medical records breakthrough that year and an order of magnitude capability gain within 24 to 36 months. Beth Israel Deaconess Medical Center will test DxO across Boston emergency and urgent care, comparing its recommendations with standard treatment for consenting patients. Google and Beth Israel published a Lancet study this month of AMIE in physician-monitored urgent care visits. Microsoft, already selling clinical dictation and radiology tools, wants decision support integrated across the full arc of care.
Positives
- MAI-DxO diagnosed 83% of complex simulated cases, compared with 36% for generalist physicians working individually with reference materials.
- DxO raised GPT-5.5 accuracy from 72% to nearly 87% on complex cases while slightly reducing estimated cost.
- DxO lifted Gemini 2.5 Flash from 46.4% to nearly 69%, indicating orchestration can substantially improve weaker models.
- Beth Israel Deaconess Medical Center will evaluate DxO alongside standard emergency and urgent care across real hospital settings.
- Microsoft and Mayo Clinic plan to release a healthcare foundation model in 2027.
Risks & concerns
- Specialists and physician teams, who usually handle cases this complex, were excluded from the 83% versus 36% comparison.
- GPT-5 already scored 78.9% alone, while DxO added only about four points despite an estimated per-case cost above $5,000.
- Emergency department tests showed little accuracy or cost improvement from DxO, including roughly 85% accuracy for both GPT-5.5 configurations.
- The performance gap between weaker and stronger foundation models exceeded the improvement delivered by Microsoft’s orchestration system.
- Real hospital evaluations must determine whether simulated diagnostic gains translate safely across diverse patients and clinical settings.