Microsoft’s Mustafa Suleyman says its AI outperforms docs on onerous instances as he chases ‘medical superintelligence’

Microsoft’s AI diagnostic system accurately recognized complicated medical instances in a simulation greater than twice as usually as generalist physicians, in keeping with a study published Thursday in Science. Mustafa Suleyman, Microsoft AI’s CEO, referred to as the consequence a step towards “medical superintelligence.”

The system, dubbed MAI-DxO (DxO for brief), coordinates a number of AI brokers that act like a panel of docs. They query a affected person, order exams, and debate a prognosis.

In a simulation utilizing complicated instances from the New England Journal of Medicine, it was appropriate 83% of the time. Primary care docs and basic internists, working individually with reference supplies, have been appropriate 36% of the time. One observe: instances this complicated are normally dealt with by specialists or doctor groups, who weren’t a part of the comparability.

“There is one thing very substantial and thrilling right here,” Dominic King, vp of Health at Microsoft AI, informed TechCrunch. “The greatest problem in healthcare supply for the time being is the rising mismatch between demand and provide of medical experience.”

In publishing these results, Microsoft joins Google and OpenAI in claiming that its personal AI can match or outperform docs.

The research’s personal knowledge means that a lot of the advance comes from extra highly effective underlying AI fashions, with Microsoft’s orchestration system including beneficial properties on prime.

Microsoft examined MAI-DxO with fashions from a number of firms, together with OpenAI, Google, Anthropic, Meta, and DeepSeek, however its finest outcomes got here with OpenAI’s GPT-5 and GPT-5.5. 

Microsoft measured DxO on price in addition to accuracy, which means the estimated price of working the AI on every case. On its personal with a brief, primary immediate, GPT-5 scored 78.9% on complicated instances at about $8,400 per case. DxO reduce that price by about $3,000 however raised accuracy by solely about 4 factors. On emergency division instances, there was little enchancment in both accuracy or price enchancment between GPT-5 by itself and with the help from DxO.

With GPT-5.5, a more recent mannequin launched in April 2026, the outcomes cut up. On complicated instances, DxO elevated GPT-5.5’s accuracy rating from 72% to almost 87% at a barely cheaper price. However, the beneficial properties as soon as once more didn’t present up on emergency division instances. Both (with and with out DxO) scored about 85% at the same price. 

DxO helps most on complicated instances and with much less succesful AI fashions. Gemini 2.5 Flash, for instance, scored 46.4% by itself however almost 69% with DxO. Even so, the hole between the weaker and stronger fashions was bigger than the enhance DxO offered.

Microsoft’s researchers stated that accuracy beneficial properties naturally shrink as fashions method the theoretical most. “The DxO orchestration helps with the broader problem of optimization each accuracy and price,” a spokesperson added.

Suleyman, who leads Microsoft’s effort to construct its personal frontier AI, informed TechCrunch that the corporate is working with Mayo Clinic to coach a basis mannequin for healthcare, with launch deliberate for subsequent 12 months. He believes this mannequin will allow the corporate to “make actually monumental leaps right here,” and says DxO is a part of the foundational transport systems Microsoft is constructing to see actual breakthroughs in healthcare. 

“I actually consider that in 2027, we will likely be on the verge of a coding type breakthrough for medical information,” Suleyman stated. “I believe we’re going to have the ability to stream extremely correct predictions about how a medical document may unfold, and clearly that form of foresight for a clinician is gold mud.”

Suleyman has spent his profession making an attempt to make use of AI to resolve medical points. He was beforehand a co-founder of DeepMind, which was acquired by Google in 2014. The following 12 months, Google DeepMind acquired King’s startup Hark, a medical activity administration app, and the 2 of them launched DeepMind Health collectively. Suleyman joined Microsoft to steer its AI unit in 2024, and King adopted a couple of months later.

“We’re simply true believers,” Suleyman stated.

He stated whereas the usage of generative AI in healthcare continues to be early, he expects healthcare fashions to get “an order of magnitude higher in functionality throughout the subsequent 24 to 36 months.” 

“As we’ve seen a number of instances within the final 5 or 6 years, the progress occurs in a short time upon getting a flywheel,” Suleyman stated. “A flywheel means now we have a mannequin in manufacturing which we’re iterating with actual globe utilization as a way to determine gaps within the mannequin after which patch them and enhance them. That’s what’s occurred in coding. It’s what occurred in chat earlier than that. And I believe subsequent 12 months it’s going to begin to occur for medical information and for medical resolution assist.”

Through a brand new partnership with Beth Israel Deaconess Medical Center in Boston, Microsoft will consider the medical reasoning capabilities of DxO throughout totally different hospital settings.

“This know-how is prepared for actual globe evaluations, and you may think about the complexity and variety of points that come from a Boston emergency division on a Friday night,” King stated.

He defined how the research will work: sufferers will get commonplace therapy, together with triage, time with a specialist, therapies, investigations, and resolution on whether or not it’s secure to discharge. In parallel, those that consent will even interact with DxO. Microsoft and its companions will examine DxO’s suggestions towards the human docs’. 

Google published a medical research alongside Beth Israel in The Lancet this month to guage Google’s AMIE conversational diagnostic AI in a real-world pressing care visits, whereas human physicians monitored the interactions in actual time. 

“We suppose this would be the greatest research of its variety,” King stated. “It will likely be throughout not simply emergency settings however pressing care environments, and can give us very compelling proof by the tip of the  research interval across the capability for these programs to be deployed at higher scale sooner or later.”

Microsoft’s research in Science coincides with a manifesto of types round what it is going to take to construct medical superintelligence, outlined as “the capability to surpass the strongest up to date human-led well being programs in medical reasoning and coordination throughout the total arc of care.” Microsoft hopes to maneuver past benchmarking particular person duties — like answering medical examination questions and producing diagnoses — into integrating with clinicians and well being care programs for higher outcomes. 

“This is the form of know-how that everyone within the globe needs, and that’s why we’re dedicated to doing every thing we are able to to speed up its introduction to the globe,” Suleyman stated, including that this focus might be a differentiator for Microsoft in comparison with the opposite frontier labs. 

Microsoft already sells dictation and radiology instruments to clinicians and sees DxO as an enlargement to these efforts. 

“We should have absolutely the primary state-of-the-art medical medical resolution assist merchandise over the following few years,” Suleyman stated.