Microsoft’s Mustafa Suleyman says its AI outperforms doctors on hard cases as he chases ‘medical superintelligence’

Microsoft’s AI diagnostic strategy correctly diagnosed analyzable aesculapian cases successful a simulation much than doubly arsenic often arsenic generalist physicians, according to a study published Thursday successful Science. Mustafa Suleyman, Microsoft AI’s CEO, called the effect a measurement toward “medical superintelligence.”

The system, dubbed MAI-DxO (DxO for short), coordinates respective AI agents that enactment similar a sheet of doctors. They question a patient, bid tests, and statement a diagnosis.

In a simulation utilizing analyzable cases from the New England Journal of Medicine, it was close 83% of the time. Primary attraction doctors and wide internists, moving individually with notation materials, were close 36% of the time. One note: cases this analyzable are usually handled by specialists oregon doc teams, who weren’t portion of the comparison.

“There is thing precise important and breathtaking here,” Dominic King, vice president of Health astatine Microsoft AI, told TechCrunch. “The biggest situation successful healthcare transportation astatine the infinitesimal is the increasing mismatch betwixt request and proviso of aesculapian expertise.”

In publishing these results, Microsoft joins Google and OpenAI successful claiming that its ain AI tin lucifer oregon outperform doctors.

The study’s ain information suggests that overmuch of the betterment comes from much almighty underlying AI models, with Microsoft’s orchestration strategy adding gains connected top.

Microsoft tested MAI-DxO with models from respective companies, including OpenAI, Google, Anthropic, Meta, and DeepSeek, but its champion results came with OpenAI’s GPT-5 and GPT-5.5. 

Accuracy improvements delivered by DxO (no fund constraints) connected CPC implicit models
prompted with baseline punctual crossed antithetic ample connection models.Image Credits:Microsoft AI

Microsoft measured DxO connected outgo arsenic good arsenic accuracy, meaning the estimated outgo of moving the AI connected each case. On its ain with a short, basal prompt, GPT-5 scored 78.9% connected analyzable cases astatine astir $8,400 per case. DxO chopped that outgo by astir $3,000 but raised accuracy by lone astir 4 points. On exigency section cases, determination was small betterment successful either accuracy oregon outgo betterment betwixt GPT-5 connected its ain and with the assistance from DxO.

With GPT-5.5, a newer exemplary released successful April 2026, the results split. On analyzable cases, DxO accrued GPT-5.5’s accuracy people from 72% to astir 87% astatine a somewhat cheaper cost. However, the gains erstwhile again didn’t amusement up connected exigency section cases. Both (with and without DxO) scored astir 85% astatine a akin cost. 

DxO helps astir connected analyzable cases and with little susceptible AI models. Gemini 2.5 Flash, for example, scored 46.4% connected its ain but astir 69% with DxO. Even so, the spread betwixt the weaker and stronger models was larger than the boost DxO provided.

Microsoft’s researchers said that accuracy gains people shrink arsenic models attack the theoretical maximum. “The DxO orchestration helps with the broader situation of optimization some accuracy and cost,” a spokesperson added.

Suleyman, who leads Microsoft’s effort to physique its ain frontier AI, told TechCrunch that the institution is moving with Mayo Clinic to bid a instauration exemplary for healthcare, with merchandise planned for adjacent year. He believes this exemplary volition alteration the institution to “make truly tremendous leaps here,” and says DxO is portion of the foundational infrastructure Microsoft is gathering to spot existent breakthroughs successful healthcare. 

“I truly judge that successful 2027, we volition beryllium connected the verge of a coding benignant breakthrough for aesculapian records,” Suleyman said. “I deliberation we’re going to beryllium capable to watercourse highly close predictions astir however a aesculapian grounds mightiness unfold, and evidently that benignant of foresight for a clinician is golden dust.”

Suleyman has spent his vocation trying to usage AI to lick aesculapian issues. He was antecedently a co-founder of DeepMind, which was acquired by Google successful 2014. The pursuing year, Google DeepMind acquired King’s startup Hark, a objective task absorption app, and the 2 of them launched DeepMind Health together. Suleyman joined Microsoft to pb its AI portion successful 2024, and King followed a fewer months later.

“We’re conscionable existent believers,” Suleyman said.

He said portion the usage of generative AI successful healthcare is inactive early, helium expects healthcare models to get “an bid of magnitude amended successful capableness wrong the adjacent 24 to 36 months.” 

“As we’ve seen aggregate times successful the past 5 oregon six years, the advancement happens precise rapidly erstwhile you person a flywheel,” Suleyman said. “A flywheel means we person a exemplary successful accumulation which we are iterating with existent satellite usage successful bid to place gaps successful the exemplary and past spot them and amended them. That’s what’s happened successful coding. It’s what happened successful chat earlier that. And I deliberation adjacent twelvemonth it’s going to commencement to hap for aesculapian records and for objective determination support.”

Through a caller concern with Beth Israel Deaconess Medical Center successful Boston, Microsoft volition measure the objective reasoning capabilities of DxO crossed antithetic infirmary settings.

“This exertion is acceptable for existent satellite evaluations, and you tin ideate the complexity and diverseness of issues that travel from a Boston exigency section connected a Friday evening,” King said.

He explained however the survey volition work: patients volition get modular treatment, including triage, clip with a specialist, treatments, investigations, and determination connected whether it’s harmless to discharge. In parallel, those who consent volition besides prosecute with DxO. Microsoft and its partners volition comparison DxO’s recommendations against the quality doctors’. 

Google published a objective survey alongside Beth Israel successful The Lancet this period to measure Google’s AMIE conversational diagnostic AI successful a real-world urgent attraction visits, portion quality physicians monitored the interactions successful existent time. 

“We deliberation this volition beryllium the biggest survey of its kind,” King said. “It volition beryllium crossed not conscionable exigency settings but urgent attraction environments, and volition springiness america precise compelling grounds by the extremity of the  survey play astir the quality for these systems to beryllium deployed astatine greater standard successful the future.”

Microsoft’s survey successful Science coincides with a manifesto of sorts astir what it volition instrumentality to physique aesculapian superintelligence, defined arsenic “the capableness to surpass the strongest modern human-led wellness systems successful aesculapian reasoning and coordination crossed the afloat arc of care.” Microsoft hopes to determination beyond benchmarking idiosyncratic tasks — similar answering aesculapian exam questions and generating diagnoses — into integrating with clinicians and wellness attraction systems for amended outcomes. 

“This is the benignant of exertion that everybody successful the satellite wants, and that’s wherefore we’re committed to doing everything we tin to accelerate its instauration to the world,” Suleyman said, adding that this absorption could beryllium a differentiator for Microsoft compared to the different frontier labs. 

Microsoft already sells dictation and radiology tools to clinicians and sees DxO arsenic an enlargement to those efforts. 

“We person to person the implicit fig 1 state-of-the-art aesculapian objective determination enactment products implicit the adjacent fewer years,” Suleyman said.