Microsoft Tests DxO in Hospitals Aiming for Medical Superintelligence

Microsoft presents DxO, an AI agent orchestration system, as capable of making the correct diagnosis in 83% of complex cases in simulation, compared to 36% for generalists. The gains vary significantly depending on the underlying model and scenarios, and a hospital study is planned with Beth Israel in Boston. Mustafa Suleyman discusses a trajectory towards "medical superintelligence" and aims for a co-trained health model with the Mayo Clinic as early as next year.

Planned hospital trials and an ambition to surpass benchmark human care

Microsoft plans to evaluate the clinical reasoning capabilities of DxO in various hospitals through a partnership with Beth Israel Deaconess Medical Center in Boston. Dominic King believes the technology is ready for real-world assessments and cites the variety of situations encountered in an emergency department on a Friday night. The announced protocol includes standard care, while consenting patients will also interact with DxO to compare its recommendations with those of doctors. This initiative accompanies a manifesto defining "medical superintelligence" as the ability to surpass the best human health systems in reasoning and coordination throughout the care pathway, with a clear intention to integrate these systems with clinicians to improve outcomes. King anticipates the largest study of its kind, covering emergency and urgent care, with evidence expected by the end of the study period for a larger-scale deployment. Mustafa Suleyman expresses a desire to accelerate the introduction of these technologies, seeing it as a potential differentiator against other labs, while Microsoft already highlights its commercial offerings in dictation and radiology tools. Meanwhile, Google published a clinical study this month in The Lancet with Beth Israel around AMIE for urgent care visits under medical supervision.

What simulations show: 83% on complex cases, limited comparison

In a simulation using complex cases drawn from the New England Journal of Medicine, DxO found the correct solution in 83% of situations. Primary care physicians and general internists, operating alone and relying on reference documents, achieved the correct diagnosis in 36% of cases. Neither specialists nor multidisciplinary teams were included in the comparison, although these cases are typically assigned to them. According to the published data, most of the gains come from the power of the underlying models, with orchestration adding additional improvements. The results were presented in Science, where it is also noted that DxO outperformed generalists more than twofold in simulation.

Cost and accuracy: orchestration primarily aids complex cases and weaker models

Microsoft measured the performance of DxO in terms of both accuracy and computational cost per case. With a short prompt, GPT-5 alone achieved 78.9% on complex cases for about $8,400 per case; when added on top, DxO reduced the cost by about $3,000 for a gain of approximately four percentage points in accuracy, with no notable benefit in emergency contexts. With GPT-5.5 released in April 2026, orchestration improved accuracy on complex cases from 72% to nearly 87% while slightly reducing costs; in emergencies, both approaches plateaued around 85% at similar costs. The most significant increases appear with less performant models: Gemini 2.5 Flash goes from 46.4% alone to nearly 69% with DxO, although the gap between weak and strong models remains more decisive than the orchestration's contribution alone. Microsoft researchers note diminishing returns as they approach a theoretical accuracy ceiling, with a spokesperson adding that DxO's coordination aims to optimize both cost and accuracy.

Which models under the hood: better scores with GPT-5 and GPT-5.5

DxO was evaluated with models from OpenAI, Google, Anthropic, Meta, and DeepSeek. The best performances were achieved with GPT-5 and especially GPT-5.5, where orchestration raised accuracy on complex cases from about 72% to nearly 87% at a slightly lower cost. In emergency situations with GPT-5.5, scores hover around 85% with or without orchestration, at comparable costs.

Aiming for a health model with the Mayo Clinic and set timelines

Mustafa Suleyman indicates that Microsoft is training, with the Mayo Clinic, a foundational model intended for health, with a release announced for next year, seeing in this component the potential for "very significant advancements" and positioning DxO within the infrastructure required for breakthroughs. He aims, within a few years, to develop some of the most advanced clinical decision support products and anticipates a tenfold increase in the capabilities of health models over 24 to 36 months. He mentions a "flywheel" where a production model is improved by real-world use, citing coding and chat as precedents, and expects that next year this cycle will begin for medical records and decision support. He also estimates that by 2027 they will be on the verge of a breakthrough in the style of coding records and believe they can provide very accurate predictions about their evolution. In this context, Dominic King points out the supply-demand imbalance of medical expertise that these tools aim to address, while Microsoft positions itself alongside Google and OpenAI among those claiming performance at or above that of physicians.

Background of the leaders and academic and clinical competition

Mustafa Suleyman, co-founder of DeepMind acquired by Google in 2014, launched DeepMind Health with Dominic King after Google's acquisition of the startup Hark the following year. He took the helm of Microsoft's AI unit in 2024, joined by King shortly after, and presents himself as a "true believer" in the contribution of AI to health. In terms of competition, Google has already conducted a clinical study with Beth Israel, published this month in The Lancet, on AMIE, its conversational diagnostic AI in urgent care contexts, under real-time physician supervision.