Microsoft's Neural Network Beats Doctors in Diagnosing Complex Diseases Microsoft introduced the MAI-DxO system: in research in Science, it correctly diagnoses 83% of complex cases, general practitioners - 36%. In typical cases - 80% compared to 47%. The system works step by step, like a doctor: several AI agents generate hypotheses, decide which tests to request, challenge assumptions and estimate the cost of a step. Imitation of counseling. According to calculations, the diagnosis is cheaper: the doctors pay about 4,725 dollars per case, the system is suitable for smaller amounts. But the figures are estimates, not from hospital bills. Disclaimers: A comparison with therapists, not experts. Test on 304 cases in the archive, not in a real hospital. And a significant part of the increase is due to the power of the base model: GPT-5 without addition gives 78.9%. Next is a prospective study with Beth Israel Deaconess in Boston, where the system will work in parallel with the doctors. A leading medical aggregator.