Microsoft’s AI diagnostic system accurately diagnosed complex medical cases in simulation more than twice as often as general physicians, according to research published Thursday in the journal Science. Mustafa Suleiman, CEO of Microsoft AI, called the results a step toward “medical superintelligence.”
The system, called MAI-DxO (DxO for short), coordinates multiple AI agents that function like a team of doctors. They ask patients questions, order tests, and discuss diagnoses.
A simulation using a complex case from the New England Journal of Medicine was correct 83% of the time. Primary care physicians and general internists worked independently with the reference materials and answered correctly 36% of the time. Note: Such complex cases are usually handled by a specialist or team of doctors who are not involved in the comparison.
“There’s something very substantial and exciting here,” Dominic King, vice president of health at Microsoft AI, told TechCrunch. “The biggest challenge in healthcare delivery at this time is the growing mismatch between the demand and supply of medical expertise.”
By announcing these results, Microsoft joins Google and OpenAI in claiming that its AI can match or outperform doctors.
The study’s own data suggests that much of the improvement is due to stronger underlying AI models, with Microsoft’s orchestration system adding benefits on top of that.
Microsoft tested MAI-DxO using models from multiple companies, including OpenAI, Google, Anthropic, Meta, and DeepSeek, and the best results were obtained with OpenAI’s GPT-5 and GPT-5.5.
Prompt using baseline prompts across a variety of large language models.Image credit: Microsoft AI
Microsoft measured DxO on cost and accuracy, or the estimated cost of running the AI in each case. GPT-5 scored 78.9% on complex cases with only short, basic prompts. The cost is approximately $8,400 per case. DxO reduced that cost by about $3,000, but improved accuracy by only about 4 points. In emergency department cases, there was little improvement in accuracy or cost between using GPT-5 alone and with the assistance of DxO.
A new model, GPT-5.5, released in April 2026 had mixed results. In complex cases, DxO improved GPT-5.5’s accuracy score from 72% to nearly 87% at a slightly lower cost. However, there was no increase again in emergency department cases. Both (with and without DxO) scored around 85% at similar costs.
DxO is most useful for complex cases and less powerful AI models. For example, Gemini 2.5 Flash scored 46.4% on its own, but scored nearly 69% on DxO. Still, the difference between the weak and strong models was greater than the boost provided by DxO.
Microsoft researchers say the accuracy gains naturally taper off as the model approaches its theoretical maximum. “DxO Orchestration helps with the broader challenge of optimizing both accuracy and cost,” the spokesperson added.
Suleiman, who is leading Microsoft’s efforts to develop its own frontier AI, told TechCrunch that the company is working with the Mayo Clinic to train a basic model for healthcare, which is expected to be released next year. He believes this model will allow the company to take “a really big leap forward here,” and says DxO is part of the foundational infrastructure Microsoft is building to enable real breakthroughs in healthcare.
“I believe that 2027 will see a breakthrough in medical records coding style,” says Suleiman. “I think we’ll be able to stream very accurate predictions about how medical records are going to play out. Obviously, that kind of foresight is gold dust for clinicians.”
Suleiman has spent his career using AI to solve medical problems. He was previously a co-founder of DeepMind, which was acquired by Google in 2014. The following year, Google DeepMind acquired King’s startup Hark, which develops clinical task management apps, and together they launched DeepMind Health. Suleiman joined Microsoft as head of AI in 2024, followed by King a few months later.
“We are just true believers,” Suleiman said.
He said that while the use of generative AI in healthcare is still in its infancy, he expects healthcare models to “become orders of magnitude more capable within the next 24 to 36 months.”
“As we’ve seen many times over the past five or six years, once you get the flywheel, progress happens very quickly,” Suleiman said. “Flywheeling means you have a model in production, and you’re iterating on real-world usage to identify gaps in the model and patch and improve it. That’s what happened with coding. It happened with chat before that. And I think next year we’ll start to see that happen with medical records and clinical decision support.”
Through a new partnership with Beth Israel Deaconess Medical Center in Boston, Microsoft will evaluate DxO’s clinical reasoning capabilities in a variety of hospital settings.
“This technology is ready for real-world evaluation, and you can imagine the complexity and variety of problems coming in from the Boston emergency department on a Friday night,” King said.
He explained how the study works. Patients will receive standard care, including triage, time with specialists, treatment, investigations, and determining whether it is safe to be discharged. In parallel, those who consent will also participate in DxO. Microsoft and its partners plan to compare DxO’s recommendations with those of human doctors.
Earlier this month, Google announced a clinical study in collaboration with Beth Israel that evaluated Google’s AMIE conversational diagnostic AI in real emergency departments, with human doctors monitoring the interaction in real time.
“We believe this is the largest study of its kind,” King said. “This applies not only to emergencies but also to critical care settings, and by the end of the study period we will have very convincing evidence as to whether these systems can be deployed at scale in the future.”
Microsoft’s research in the journal Science aligns with a manifesto of sorts for what it takes to build a medical superintelligence. This is defined as “the ability to exceed the strongest human-driven modern medical systems in medical reasoning and coordination across health care.” Microsoft wants to move beyond benchmarking individual tasks, such as answering health exam questions or generating diagnoses, to integrating with clinicians and health systems for better results.
“This is the kind of technology that everyone in the world wants, so we’re committed to doing everything we can to accelerate its adoption into the world,” Suleiman said, adding that this focus could be a differentiator for Microsoft compared to other Frontier Labs.
Microsoft already sells dictation and radiology tools to clinicians, and sees DxO as an extension of those efforts.
“In the coming years, we must have the absolute number one, most advanced medical clinical decision support product,” Suleiman said.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
