Lead (5W1H) What: An AI‑driven diagnostic platform called MAI‑DxO correctly diagnosed complex medical cases at a rate more than double that of generalist physicians. Who: Microsoft AI, led by CEO Mustafa Suleyman, conducted the study. When: The results were published on Thursday in the journal Science. Where: The evaluation used case studies originally featured in the New England Journal of Medicine. Why: Microsoft aims to develop what it calls “medical superintelligence” to address the growing gap between demand for and supply of medical expertise. How: MAI‑DxO orchestrates multiple AI agents that interview patients, order tests, and debate diagnoses, mimicking a multidisciplinary medical panel.
Study results
According to TechCrunch, the Science paper reported that MAI‑DxO achieved an 83 % correctness rate on the complex case set, while primary‑care doctors and general internists working alone with reference materials were correct 36 % of the time. The article notes that the comparison excluded specialist physicians or multidisciplinary teams, who normally handle such difficult cases.
Dominic King, Microsoft AI’s vice‑president of Health, told TechCrunch that the findings highlight “the biggest challenge in healthcare delivery at the moment is the growing mismatch between demand and supply of medical expertise.”
Technical underpinnings
TechCrunch explains that the performance boost stems from two sources: the underlying large language models (LLMs) and the DxO orchestration layer. Microsoft tested MAI‑DxO with models from OpenAI, Google, Anthropic, Meta and DeepSeek. The best outcomes were recorded with OpenAI’s GPT‑5 and GPT‑5.5.
When run with a simple prompt, GPT‑5 alone scored 78.9 % accuracy on the complex cases at an estimated $8,400 per case. Adding DxO reduced the cost by roughly $3,000 while nudging accuracy upward by about four percentage points.
For emergency‑department scenarios, TechCrunch reports that neither accuracy nor cost improved appreciably with DxO; both the baseline GPT‑5 and the DxO‑augmented version hovered around 85 % accuracy at similar expense.
With the newer GPT‑5.5 model, DxO raised accuracy from 72 % to nearly 87 % on complex cases and achieved a modest cost reduction. Again, emergency‑department performance remained flat at about 85 %.
The article adds that DxO’s greatest impact appears on “complex cases and with less capable AI models.” For example, Gemini 2.5 Flash rose from 46.4 % to 69 % accuracy when paired with DxO. However, TechCrunch notes that the gap between weaker and stronger models still exceeds the incremental gains provided by the orchestration layer.
Microsoft’s researchers, as reported by TechCrunch, argue that accuracy improvements naturally taper as models near their theoretical maximum, and that DxO’s value lies in jointly optimizing accuracy and cost.
Roadmap toward “medical superintelligence”
TechCrunch cites Suleyman’s statement that Microsoft is collaborating with the Mayo Clinic to train a dedicated healthcare foundation model, slated for release next year. Suleyman envisions a “coding‑style breakthrough for medical records” by 2027, enabling real‑time, highly accurate predictions of how a patient’s record might evolve.
Suleyman’s background, according to TechCrunch, includes co‑founding DeepMind (acquired by Google in 2014) and launching DeepMind Health after acquiring the clinical‑task app Hark. He joined Microsoft’s AI unit in 2024, with King following shortly thereafter.
Both executives, as quoted by TechCrunch, expect healthcare‑focused AI models to become “an order of magnitude better in capability within the next 24 to 36 months,” driven by a “flywheel” of production deployment, real‑world feedback, and iterative improvement—paralleling progress seen in coding and chat AI.
Real‑world evaluation partnership
TechCrunch reports that Microsoft has partnered with Beth Israel Deaconess Medical Center in Boston to test DxO’s clinical reasoning across multiple hospital settings. In the planned study, patients will receive standard care while consenting participants also interact with DxO; the AI’s recommendations will then be compared to those of human clinicians.
The article mentions that Google recently published a Lancet study with Beth Israel evaluating its AMIE conversational diagnostic AI in urgent‑care visits, suggesting a competitive landscape. King, as reported by TechCrunch, describes Microsoft’s upcoming trial as “the biggest study of its kind,” spanning emergency and urgent‑care environments and aiming to provide compelling evidence for large‑scale deployment.
Strategic positioning
According to TechCrunch, Microsoft’s Science paper includes a manifesto defining “medical superintelligence” as the ability to surpass the strongest contemporary human‑led health systems in reasoning and coordination across the full arc of care. The company seeks to move beyond isolated benchmarks (e.g., exam‑question answering) toward deeper integration with clinicians and health‑system workflows.
Suleyman, as quoted by TechCrunch, argues that this focus could differentiate Microsoft from other frontier AI labs, noting that the technology “is the kind of technology that everybody in the world wants.” Microsoft already markets dictation and radiology tools to clinicians, and DxO is positioned as an expansion of that portfolio. Suleyman asserts that Microsoft aims to become the “absolute number one state‑of‑the‑art medical clinical decision‑support product” provider in the coming years.
Brief analysis (opinion) The reported performance gap between MAI‑DxO and generalist physicians is striking, yet the study’s scope—simulation using published case reports and a comparison limited to individual primary‑care doctors—means the results may not directly translate to real‑world hospital settings. The upcoming Beth Israel trial will be critical for validating whether the orchestration approach can sustain its gains amid the variability of live patient encounters.
Conclusion Microsoft’s MAI‑DxO system demonstrates notable accuracy improvements over generalist physicians on simulated complex cases, especially when paired with leading LLMs like OpenAI’s GPT‑5.5. While cost reductions and performance lifts are modest, the technology underscores Microsoft’s ambition to achieve “medical superintelligence” and to embed AI more deeply within clinical decision‑support workflows. Real‑world trials slated for 2027 will be decisive in determining the platform’s practical impact.
Source: TechCrunch
Frequently Asked Questions
What was the primary metric used to compare MAI‑DxO with physicians?**
Accuracy of correct diagnoses on a set of complex cases taken from the *New England Journal of Medicine*, with MAI‑DxO achieving 83 % and generalist physicians 36 %, according to TechCrunch. **Q: Which underlying AI models performed best with DxO?** A: OpenAI’s GPT‑5 and GPT‑5.5 delivered the highest accuracy scores when combined with DxO, as reported by TechCrunch. **Q: How will Microsoft test DxO in a clinical environment?** A: Through a partnership with Beth Israel Deaconess Medical Center, Microsoft will compare DxO’s recommendations to standard care for consenting patients across emergency and urgent‑care settings, according to TechCrunch.