Microsoft AI published a study in Science on October 22, 2026 reporting that its diagnostic system, MAI-DxO, correctly diagnosed complex cases in a simulation more than twice as often as generalist physicians. According to TechCrunch's report on the paper, the system scored 83% on cases drawn from the New England Journal of Medicine, against 36% for doctors working alone. Microsoft AI CEO Mustafa Suleyman described the work as a step toward "medical superintelligence."
The headline is easy to repeat. The more useful reading is in the details the study itself surfaces: how much of the gain comes from the orchestrator versus the underlying model, what the comparison group was, and what a simulation cannot tell us.
What MAI-DxO is
MAI-DxO coordinates several AI agents that behave like a panel of doctors. In the simulation the agents question a simulated patient, order tests and debate a diagnosis before committing to one, per TechCrunch. It is an orchestration layer rather than a new foundation model: the study tested models from OpenAI, Google, Anthropic, Meta and DeepSeek inside the same framework.
Microsoft
OpenAI
Google
AnthropicThe results, side by side
Diagnostic accuracy on complex NEJM cases
Share of cases diagnosed correctly in a simulation. Doctors were primary care physicians and general internists working alone. Values marked approximate are reported as 'about' or 'nearly'.
Source: TechCrunch report on the Microsoft AI study in Science, October 22, 2026. Approximate: Gemini 2.5 Flash with DxO (about 69%) and GPT-5.5 with DxO (nearly 87%).
| Configuration | Complex-case accuracy |
|---|---|
| Doctors working alone | 36% |
| Gemini 2.5 Flash alone | 46.4% |
| Gemini 2.5 Flash with MAI-DxO | About 69% |
| GPT-5.5 alone | 72% |
| GPT-5 alone (short, basic prompt) | 78.9% |
| MAI-DxO (no budget constraint) | 83% |
| GPT-5.5 with MAI-DxO | Nearly 87% |
Two patterns stand out. First, a strong base model with a simple prompt already beat the doctors by a wide margin: GPT-5 alone scored 78.9%. Second, the orchestrator helped most where the base model was weaker, lifting Gemini 2.5 Flash by roughly 23 points (our arithmetic from the reported 46.4% and about 69%), while adding less on top of the strongest models.
Analysis: how much is the model and how much is the system?
TechCrunch notes that the study's own data suggests most of the gain comes from the underlying AI models, with MAI-DxO adding improvements on top. Microsoft's researchers also say accuracy gains shrink as models approach the theoretical maximum. For a reader trying to judge the result, this changes the question from "does an orchestrator beat doctors" to three narrower ones:
| Question | What the reported data says |
|---|---|
| Does model choice matter more than orchestration? | The gap between weaker and stronger models was larger than the orchestrator's gain, per TechCrunch |
| Does orchestration still add value? | Yes, most visibly for weaker models and complex cases |
| Is it uniformly useful? | No. Emergency department cases showed little benefit; GPT-5.5 was about 85% alone and with the system |
The cost angle
TechCrunch reports that GPT-5 alone cost about $8,400 per case in the study's setup, and that MAI-DxO reduced that by about $3,000 while raising accuracy only about four points. With GPT-5.5, the orchestrated system cost slightly less than the base model, though the article gives no dollar figure. These are simulation costs for model inference and ordered tests as the study modelled them, not clinical prices, and they should not be read as the cost of care.
What the study cannot tell us
- It is a simulation. The cases came from the New England Journal of Medicine and the agents interrogated a simulated patient. Real patients present incomplete histories, messy records and time pressure.
- The comparison group was narrow. Primary care doctors and general internists worked individually with reference materials. Specialists and physician teams, who usually handle cases this complex, were not part of the comparison.
- Emergency care showed little gain. The system's advantage did not carry over to emergency department cases in the reported results.
- The result partly measures the models. The best results used GPT-5 and GPT-5.5, so the headline reflects those models as much as the orchestration.
The honest position is that the study shows the ceiling of a method, not the performance of a clinic.
Metir analysis
What happens next
Per TechCrunch, Microsoft is partnering with Beth Israel Deaconess Medical Center in Boston to evaluate MAI-DxO's reasoning in emergency and urgent care. Patients who consent will receive its input alongside standard care, and its recommendations will be compared with those of human doctors. Microsoft is also working with the Mayo Clinic on a healthcare foundation model planned for 2027. Google published a similar study of its AMIE system in The Lancet with Beth Israel, and OpenAI and Google have also claimed their AI can match or outperform doctors, so this is a contested category with several parallel evidence tracks.
Why it matters beyond medicine
The result is a clean example of a pattern that applies well outside healthcare: an orchestration layer's value depends on the model underneath it, and the underlying model changes every few months. A system that was best with GPT-5 may rank differently with GPT-5.5 or a Gemini or Claude model. Teams that build agent workflows on a single provider carry that dependency. Tools that let you swap models behind a stable workflow, as Metir does for chat, make that kind of comparison a configuration change rather than a rebuild.
What to watch
- Real-world results from Beth Israel. Consent-based clinical use is the first test outside the simulation.
- Specialist and team baselines. A comparison with specialists would show where the 36% figure sits relative to the people who normally handle these cases.
- Replication on other models. Independent runs with newer Claude, Gemini or open-weight models would show how portable the orchestrator's gains are.
- Regulatory path. Microsoft already sells dictation and radiology tools to clinicians and describes MAI-DxO as an extension of them, which will shape how any clinical claim is reviewed.
Sources:
Image credits
- Hero and in-body photo: "Aerial Microsoft West Campus August 2009.jpg" by Jelson25, public domain, via Wikimedia Commons. Pulled 2026-10-23 and reviewed visually before use.