Mustafa Suleyman of Microsoft Asserts AI Surpasses Doctors in Complex Cases, Pursuing ‘Medical Superintelligence’
A recent publication in Science highlights that Microsoft’s AI diagnostic tool, MAI-DxO, has outperformed general practitioners in diagnosing complex medical cases during simulations, achieving success rates over twice that of human doctors. Mustafa Suleyman, CEO of Microsoft AI, termed these results as a groundbreaking step toward achieving “medical superintelligence.”
The AI solution, known as MAI-DxO (Diagnostic Operating System), operates through multiple AI agents collaborating to simulate a panel of medical experts. These agents collect patient information, conduct diagnostics, and discuss possible diagnoses.
During simulations involving complex cases acquired from the New England Journal of Medicine, MAI-DxO demonstrated an impressive diagnostic accuracy rate of 83%. In contrast, primary care physicians and general internists managed only a 36% accuracy rate, even when using reference materials. It’s important to note that the intricate cases typically require specialist teams, which were not included in this evaluation.
Dominic King, the vice president of Health at Microsoft, commented to TechCrunch that “this is highly significant and exhilarating.” He stressed the foremost challenge in healthcare—the growing divide between the demand for medical expertise and its availability.
With the release of these findings, Microsoft places itself among top firms like Google and OpenAI, asserting that their AI technologies could match or even exceed the capabilities of human physicians in certain situations.
The results suggest that this improved performance is largely due to advanced AI models combined with Microsoft’s orchestration system.
MAI-DxO was tested against various models from companies such as OpenAI, Google, Anthropic, Meta, and DeepSeek, with optimal performance achieved using OpenAI’s GPT-5 and GPT-5.5 models.

In addition to prioritizing accuracy, Microsoft assessed the cost-effectiveness of DxO in relation to AI usage for each medical case. With a standard prompt, GPT-5 achieved a 78.9% accuracy rate for complex cases at an estimated cost of roughly $8,400 per case. DxO reduced this cost by nearly $3,000 while providing a slight accuracy boost of around four percentage points. For emergency situations, no significant differences in accuracy or costs were observed between GPT-5 and DxO.
Testing GPT-5.5, a newer model introduced in April 2026, yielded mixed outcomes. DxO enhanced GPT-5.5’s accuracy from 72% to approximately 87%, while also lowering costs for complex cases. However, similar to GPT-5, no major improvements were noted in emergency department scenarios, as both models maintained around 85% accuracy at comparable costs.
DxO shows considerable skill in handling complex cases and works effectively with less capable AI models. For example, while Gemini 2.5 Flash alone achieved only 46.4% accuracy, its performance improved to nearly 69% when paired with DxO. Nevertheless, the accuracy gap between weaker and stronger models remains significant, overshadowing the improvements facilitated by DxO.
Microsoft researchers have acknowledged that as models approach their theoretical performance limits, achieving greater accuracy enhancements will likely pose challenges. “The DxO orchestration aims to address the broader challenge of aligning accuracy with costs,” stated a company spokesperson.
Suleyman, who oversees Microsoft’s advanced AI projects, disclosed to TechCrunch that they are collaborating with the Mayo Clinic to create a foundational healthcare model expected to launch next year. He is optimistic that this model will lead to “substantial advancements” and views DxO as a crucial element of the foundational architecture Microsoft is developing to drive significant improvements in healthcare.
“I firmly believe that by 2027, we will witness a major transformation in the structure of medical records,” Suleyman predicted. “We will be able to generate highly accurate forecasts about the future of a medical record, providing invaluable insights to healthcare providers.”
Suleyman has devoted his career to harnessing AI solutions to tackle medical challenges, having co-founded DeepMind, which was acquired by Google in 2014. The following year, Google DeepMind merged with King’s startup Hark, a clinical task management platform, resulting in the establishment of DeepMind Health. Suleyman joined Microsoft in 2024 to lead its AI division, soon followed by King.
“We are genuinely committed to this vision,” Suleyman emphasized.
While the application of generative AI in healthcare is still maturing, he anticipates “remarkable improvements in capability within the next 24 to 36 months.”
“As we’ve consistently seen over the past five to six years, progress can accelerate dramatically once the flywheel effect takes root,” Suleyman observed. “A flywheel means we have a functional model that we continuously refine based on real-world applications to identify and rectify weaknesses. This effect has been apparent in coding and previously in conversational technologies. I expect the upcoming year will bring similar breakthroughs in medical records and clinical decision-making support.”
Through a recent partnership with Beth Israel Deaconess Medical Center in Boston, Microsoft intends to evaluate DxO’s clinical reasoning skills across various hospital environments.
“This technology is ready for real-world testing, and one can only imagine the complexities and hurdles presented by a busy Boston emergency department on a Friday night,” King explained.
He detailed the study’s approach: patients will receive standard treatments, including triage, specialist consultations, necessary interventions, and discharge safety assessments. Concurrently, participating patients will engage with DxO, allowing Microsoft and its partners to compare AI-generated recommendations with those of human doctors.
In a similar endeavor, Google recently executed a clinical study published in The Lancet with Beth Israel to evaluate its AMIE conversational diagnostic AI during genuine urgent care visits, supervised in real-time by human physicians.
“We expect this to be the largest study of its kind,” King proclaimed. “It will encompass not just emergency situations but also urgent care settings, providing compelling evidence for the future scalability of these systems.”
The study reported by Microsoft in Science aligns with a manifesto outlining the essential criteria for achieving medical superintelligence: the capacity to surpass current human-operated healthcare systems in medical reasoning and coordination throughout comprehensive care. Microsoft seeks to go beyond merely assessing isolated tasks, such as answering medical exam questions and making diagnoses, by closely collaborating with clinicians and healthcare systems to enhance overall health outcomes.
“This is the type of technology that people globally are looking for, and that’s why we are committed to accelerating its global rollout,” Suleyman concluded, suggesting that this ambition could set Microsoft apart from other leading laboratories.
Microsoft has previously provided dictation and radiology solutions to healthcare providers and regards DxO as a natural extension of these efforts.
“We must deliver the most advanced clinical decision support tools in the coming years,” Suleyman asserted.
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.


