Microsoft’s MAI-DxO system scored 83.2% on one historical case benchmark and 80.3% on another, according to a study published October 22, 2026. Both tests used controlled, text-based simulations. The scores measure performance on those benchmarks, not outcomes for patients receiving care.
What MAI-DxO scored on each benchmark
The study evaluated MAI-DxO with GPT-5 on 304 New England Journal of Medicine clinicopathological conference cases and 325 emergency-department-to-admission encounters from MIMIC-IV-ED. Participating physicians scored lower on both benchmarks, but the comparisons involved individual physicians rather than specialist teams or clinicians working with patients.
| Benchmark cohort | Historical cases | MAI-DxO with GPT-5 | Participating physicians | Physician comparison condition |
| New England Journal of Medicine clinicopathological conferences | 304 | 83.2% | 36.1% with references; 22.3% without | Participating generalists |
| MIMIC-IV-ED emergency-department-to-admission encounters | 325 | 80.3% | 46.8% | Participating physicians |
The paper also reports a separate evaluation using GPT-5.5, with the same prompts and orchestration. GPT-5 powered the study’s principal reported results above.
How SDBench turns records into an interactive test
SDBench, the Sequential Diagnosis Benchmark, lets the system request information in stages. It starts with a short case description, then can ask for history or examination findings, order tests, revise its possible diagnoses and submit a final diagnosis.
A Gatekeeper supplies information from the historical case record in response to each request. If the record lacks a requested finding, the Gatekeeper can generate a diagnostically neutral synthetic result. A Judge then scores the final answer using a physician-authored, five-level rubric; scores of 4 or 5 count as correct.
What the accuracy and cost figures mean
SDBench replayed historical cases in a text-only setting. It did not include patient interaction, physical examination, imaging, waveforms or pathology slides. The study’s cases were pathological rather than a representative mix of healthy, benign and serious encounters, so the results do not establish false-positive rates or how the system would perform across ordinary case prevalence. The authors also say the benchmark does not establish performance against specialists, multidisciplinary teams or real-world clinical practice, and it measured neither treatment safety nor patient outcomes.
The study reported different accuracy and estimated cost combinations for the New England Journal of Medicine cases. MAI-DxO reached 83.2% accuracy at an estimated $5,431 per case without a budget limit. A budgeted configuration reached 79.6% at $1,303 per case, while an ensemble configuration reached 84.5% at $8,945 per case. These are modeled benchmark estimates, not patient bills or demonstrated savings.
The cost model mapped requested tests to billing codes and used prices from a 2023 Kaiser Permanente Washington chargemaster. It also assigned a fixed $300 cost to a physician visit. The study’s authors note that the estimates do not account for differences among providers and insurers, test availability, time to results, invasiveness or patient discomfort. The study authors were Microsoft employees and disclosed equity interests in the company.
The study describes holding out the 56 most recent New England Journal of Medicine cases from system development; most were published after the evaluated models’ training cutoffs.
Microsoft’s planned hospital evaluation
In its October 22, 2026 announcement, Microsoft said it planned a prospective evaluation with Beth Israel Deaconess Medical Center. The proposed study would assess clinical reasoning in hospital settings, with patients interacting with the system and its performance compared with physicians’ independent assessments.
