On emergency cases, GPT-5 alone scored 76.6%; inside Microsoft's system, 80.3%

Sources

• Microsoft AI, Building Medical Superintelligence, October 22, 2026 (opens in a new tab)Figure 2 table, Emergency Department cases (Models evaluated October 2025): GPT-5, MAI-DxO (no budget), Clinicians, open book (average)
• Microsoft, MIMIC-IV-Ext-SDBench (PhysioNet, version 1.0.0) (opens in a new tab)Usage Notes: the baseline agent and the Diagnostic Orchestrator with GPT-5 on the 325 validation and test cases
• Nori and others, Sequential diagnosis with language models, Science, October 22, 2026 (opens in a new tab)Abstract: generalist physicians on MIMIC-IV-ED emergency department admissions

Method

Each case starts from a short summary of an emergency department visit that ended in a hospital admission, taken from MIMIC-IV, a database of de-identified hospital records used for research. The doctor or AI asks questions and orders tests, an AI model answers from the patient's discharge summary, and another AI model grades the final diagnosis. The two GPT-5 figures cover the 325 cases Microsoft reports results on. The physicians' figure is the average for clinicians working with reference materials, from Microsoft's results table and the study's abstract, neither of which says how many of the cases the physicians worked. In the same dataset notes, the average estimated cost of the tests ordered was about $4,240 a case for GPT-5 alone and $4,980 inside MAI-DxO.

Published

November 5, 2026

Free to share and embed with credit to Kent Applied (CC BY 4.0). The source data keeps its own terms.