Microsoft GPT-5 Diagnostic System Outperforms Doctors in Science Study

A diagnostic system from Microsoft researchers, powered by GPT-5, outperformed participating physicians while often lowering estimated testing costs, in a study published in Science, according to AAAS. Harsha Nori and colleagues built Diagnostic Orchestrator, or DxO, to select questions and medical tests as new patient information becomes available, first presented by Microsoft in 2025.

The team evaluated DxO with SDBench, a benchmark that converts 629 clinical cases into step by step diagnostic tasks. The cases came from The New England Journal of Medicine and emergency department encounters. DxO, powered by a model from OpenAI, achieved accuracy from 79.6 percent to 84.5 percent, with average estimated costs from $1,303 to $8,945 per case depending on budget limits.

Physicians evaluating curated clinical conference cases averaged 22.3 percent accuracy without reference tools and 36.1 percent with them. The best performing physician reached 51.8 percent accuracy. On routine emergency department cases, DxO reached 80.3 percent accuracy at an average estimated cost of $4,980 per case without budget limits.

DxO gathered more patient history before ordering its first test than participating physicians. The researchers said further clinical evaluation is needed because the system cannot perform physical examinations, observe patients directly, or establish its effect on patient outcomes.