Lead: Microsoft’s AI diagnostic platform, MAI‑DxO, correctly identified complex medical cases in a simulation 83% of the time—more than twice the 36% accuracy recorded for primary‑care physicians—according to a study published Thursday in Science and reported by TechCrunch.
Study results
According to TechCrunch, the study compared the performance of MAI‑DxO, an AI system that coordinates multiple agents to act like a panel of doctors, against individual primary‑care doctors and general internists handling the same cases. The AI achieved an 83% correct‑diagnosis rate on simulated cases drawn from the New England Journal of Medicine, while the physicians achieved 36% when working alone with reference materials.
The article notes that the cases used in the simulation are typically managed by specialist teams, a factor not included in the physician comparison. Dominic King, vice president of Health at Microsoft AI, told TechCrunch that the findings highlight “something very substantial and exciting” amid a growing mismatch between demand for medical expertise and its supply.
How MAI‑DxO works
TechCrunch reports that MAI‑DxO orchestrates several AI agents that interview a virtual patient, order diagnostic tests, and debate possible diagnoses, mimicking a multidisciplinary clinical team. The system was evaluated with large language models (LLMs) from multiple vendors, including OpenAI, Google, Anthropic, Meta, and DeepSeek.
According to the report, the best results were obtained using OpenAI’s GPT‑5 and GPT‑5.5. When GPT‑5 was run with a basic prompt, it scored 78.9% accuracy on the complex cases at an estimated cost of $8,400 per case. Adding the DxO orchestration reduced the cost by roughly $3,000 while raising accuracy by about four percentage points.
With GPT‑5.5, the DxO layer boosted accuracy from 72% to nearly 87% on the same complex cases, with a modest cost reduction. However, the report states that neither GPT‑5 nor GPT‑5.5 showed significant accuracy or cost improvements on emergency‑department cases, where both AI‑only and AI‑plus‑DxO configurations scored around 85% at similar expense.
The article highlights that DxO’s impact is most pronounced on difficult cases and with less capable base models. For example, Gemini 2.5 Flash, a weaker model, rose from 46.4% to almost 69% accuracy when paired with DxO. Nonetheless, the gap between the strongest and weakest models remained larger than the incremental gain provided by DxO.
Cost and efficiency considerations
TechCrunch notes that Microsoft measured both accuracy and cost, emphasizing that DxO “helps with the broader challenge of optimization both accuracy and cost.” A Microsoft spokesperson, as cited by the article, added that accuracy gains naturally diminish as models approach their theoretical maximum performance.
Future roadmap
Suleyman, Microsoft’s AI chief, told TechCrunch that the company is collaborating with the Mayo Clinic to train a dedicated healthcare foundation model slated for release next year. He described this effort as a step toward “medical superintelligence,” which the article defines as the ability to surpass the strongest contemporary human‑led health systems in medical reasoning and coordination across the full arc of care.
Suleyman also projected that by 2027 the company expects a “coding‑style breakthrough for medical records,” enabling highly accurate predictions about how a record might evolve—information he called “gold dust” for clinicians. He linked this vision to a broader expectation that healthcare AI models will improve “an order of magnitude” within the next 24‑36 months, citing the rapid progress observed in other AI domains such as coding and conversational chat.
Real‑world testing
According to TechCrunch, Microsoft has partnered with Beth Israel Deaconess Medical Center in Boston to evaluate DxO in live hospital settings. The study will run parallel tracks: patients receive standard care—including triage, specialist consultation, investigations, and discharge decisions—while consenting participants also interact with DxO. Researchers will compare DxO’s recommendations with those of the human care team.
The article mentions that Google recently published a similar real‑world evaluation of its AMIE conversational diagnostic AI in The Lancet, also in partnership with Beth Israel. King told TechCrunch that Microsoft’s upcoming study “will be the biggest of its kind,” spanning emergency and urgent‑care environments and providing “very compelling evidence” about the scalability of such systems.
Strategic positioning
TechCrunch reports that Microsoft frames the Science study as part of a manifesto outlining the requirements for building medical superintelligence. The company aims to move beyond isolated benchmark tasks—such as answering exam questions or generating single diagnoses—toward integrated decision‑support tools that work alongside clinicians and health systems.
Suleyman, as quoted in the article, stated that this focus could differentiate Microsoft from other frontier AI labs, noting the company already sells dictation and radiology tools to clinicians. He emphasized the ambition to become “the absolute number one state‑of‑the‑art medical clinical decision support product” in the coming years.
Brief analysis (opinion)
The findings suggest that coordinated AI orchestration can substantially narrow the performance gap between current generalist physicians and cutting‑edge language models on highly complex cases. However, the limited improvement on emergency‑department scenarios indicates that real‑time, high‑stakes environments may require further model refinement or different orchestration strategies before widespread clinical adoption.
Conclusion
The Science paper highlighted by TechCrunch demonstrates that Microsoft’s MAI‑DxO AI system can outperform individual generalist physicians on simulated complex medical cases, achieving 83% diagnostic accuracy. While cost reductions and accuracy gains are modest, the study underscores the potential of AI orchestration, especially when paired with powerful language models. Ongoing real‑world evaluations at Beth Israel Deaconess will test whether these laboratory results translate into tangible clinical benefits, marking a critical step toward the company’s broader goal of medical superintelligence.
Source: TechCrunch