Microsoft reported in a Science paper that its MAI‑DxO (DxO) diagnostic system correctly diagnosed complex, simulated cases from the New England Journal of Medicine 83% of the time—more than twice the accuracy of individual primary care physicians and general internists working with reference materials, who were correct 36% of the time. The company published the results on a Thursday.
How DxO works
MAI‑DxO is an orchestration layer that coordinates multiple AI agents so they operate like a panel of doctors: they question a patient, order tests and debate diagnoses. Microsoft says this orchestration improves outcomes and cost‑efficiency on top of the underlying large language models.
Accuracy and cost details from the study
- DxO achieved 83% accuracy on the most complex cases drawn from the New England Journal of Medicine; individual generalist physicians scored 36% in the same simulated tasks.
- Using a short, basic prompt, OpenAI’s GPT‑5 by itself scored 78.9% on complex cases at a cost of roughly $8,400 per case. Adding DxO reduced that cost by about $3,000 and raised accuracy by about four percentage points for that configuration.
- With GPT‑5.5 (a newer model the study notes was released in April 2026), results diverged: for complex cases DxO raised GPT‑5.5’s accuracy from 72% to nearly 87% while slightly lowering cost. For emergency department cases, however, there was little difference—both with and without DxO scored about 85% at similar costs.
- DxO provided the greatest benefit on complex cases and when paired with less capable base models. For example, Gemini 2.5 Flash scored 46.4% alone and nearly 69% with DxO. Microsoft’s researchers note, though, that the performance gap between weaker and stronger models remains larger than the uplift from orchestration.
The authors also observe that accuracy gains naturally diminish as models near theoretical maxima. Microsoft describes DxO’s role as helping optimize the tradeoff between accuracy and cost.
Models tested and partnerships
Microsoft tested DxO with models from several vendors, including OpenAI, Google, Anthropic, Meta and DeepSeek; the best results were achieved with OpenAI’s GPT‑5 and GPT‑5.5. Mustafa Suleyman, CEO of Microsoft AI, characterized the outcome as a step toward “medical superintelligence.”
The company said it is working with the Mayo Clinic to train a healthcare foundation model planned for release next year, and it has a partnership with Beth Israel Deaconess Medical Center to evaluate DxO’s clinical reasoning in different hospital settings. In those planned trials, patients will receive standard care (triage, specialist review, investigations and disposition), and consenting participants will also interact with DxO in parallel so researchers can compare the system’s recommendations with clinicians’.
Context with other industry efforts and future expectations
Microsoft’s publication follows similar claims from other leading labs—Google and OpenAI—about AI systems that can match or exceed clinicians on certain tasks. Google and Beth Israel recently published a clinical study in The Lancet evaluating Google’s AMIE conversational diagnostic AI in real urgent care visits.
Dominic King, vice president of Health at Microsoft AI, told TechCrunch that the largest current challenge in healthcare delivery is the mismatch between demand and available medical expertise, and he suggested systems like DxO could help address that gap. Microsoft aims to move beyond task‑level benchmarks toward integrating AI with clinicians and healthcare systems to improve outcomes.
Mustafa Suleyman said generative AI in healthcare is still early, but he expects medical models to become “an order of magnitude better” within 24–36 months. He predicted a 2027 inflection for electronic medical records and clinical decision support, driven by production models that are iteratively improved through real‑world use. Suleyman, a former DeepMind co‑founder who joined Microsoft in 2024, and Dominic King, who joined Microsoft soon after, described their long‑term commitment to medical AI.
Next steps
Microsoft plans larger real‑world evaluations to test whether these systems can be deployed at scale. King said the planned Boston study will cover not only emergency departments but also urgent care settings, and he characterized it as potentially the biggest study of its kind. The company already sells dictation and radiology tools to clinicians and sees DxO as an extension of those efforts. Suleyman emphasized Microsoft’s goal to develop state‑of‑the‑art clinical decision support products over the coming years.
