Many MindedA panel of simulated doctors beat human physicians at hard diagnostic puzzles
Microsoft researchers have described a diagnostic system called MAI-DxO in a paper published in Science in November 2026. It runs on GPT-5 and acts like a small panel of physicians. One role proposes diagnoses. Another suggests questions and tests and challenges the leading idea. A third weighs costs. An orchestrator coordinates them and decides whether a test is worth its price.
The team built the benchmark from 304 complex New England Journal of Medicine clinicopathological cases, plus a similar number of typical emergency room cases. Each case became an interactive task that starts with limited information. Participants could ask about history, order tests and revise their hunches before committing to a diagnosis. Forty physicians, each with about 12 years of experience, served as the comparison group.
On the complex cases MAI-DxO was right 83 percent of the time, against 22 percent for physicians relying on their own knowledge. On typical emergency cases it scored 80 percent against 47 percent. A cost-conscious configuration kept nearly 80 percent accuracy at an estimated $1,300 per case, using 2023 hospital pricing, while the physicians' tests cost over $4,700. In a scurvy case, it ordered a targeted vitamin C test, while standalone GPT-5 ordered a workup costing nearly 7.6 times as much. It also beat general models from Anthropic, Google and xAI, and swapping in GPT-5.5 raised accuracy further.
The abundance case is that expert diagnostic reasoning is scarce and expensive, and this approach aims to supply it more cheaply while trimming unnecessary testing. Good diagnosis that costs less could reach more people.
The caveats are large. These were controlled, text-based simulations, not clinics. They left out unreliable histories, medical images, other clinical data and time pressure. Chronic or evolving diseases such as cancer could stump the system, and patients may decline recommended tests. The article says the results do not mean unsupervised AI diagnosticians are coming to hospitals soon. Real-world trials are the next thing to watch.
Source: Singularity Hub