Back
Medical Index
The micro1 Medical Index brings our medical benchmarks together in one place.
The rankings are a live scoreboard and will be updated with every new benchmark we publish in this domain. The Index score is the equal-weight average of the three benchmark scores, and only models evaluated on all three are ranked.
While these benchmarks capture important dimensions of medical performance, they do not represent the full scope of clinical practice. The micro1 Medical Index should therefore be interpreted as a measure of performance across the evaluated tasks, rather than a comprehensive assessment of medical competence or readiness for clinical deployment.
No confidence intervals have been computed for the aggregate score, so differences of a few thousandths should be treated as ties.
%20(1).webp)
Overview
The micro1 Medical Index premise comes from what we see across our research: frontier models now carry broad medical knowledge, so the open questions have moved to clinical judgment. That means recognizing the limits of the evidence, keeping every number traceable to its source, and reading what a question is really asking.Every benchmark in the Index is built with practicing clinicians and scientists from their own work. The aim is to measure how close model output comes to what a clinician would sign off on, and where expert review still adds the most.
The benchmarks
RealmPathology Benchmark
RealmPathology Benchmark tests whether a model can interpret a real anatomic pathology report without saying more than the report supports: reaching a diagnosis, stage or management call only when the specimen allows it. 1,071 rubric criteria, 13 frontier models.
PharmaBench
PharmaBench tests whether an AI agent can do the documentation work behind a clinical trial, inside complete synthetic drug programs, with every number right and every claim backed by the study data. 120 tasks across 8 drug programs, written by biostatisticians, medical writers, safety physicians and regulatory specialists, 10 frontier models.
WellnessBench
WellnessBench tests whether a model can respond to a young person asking about their health in a way that is both safe and still useful, including when the user pushes back. A dataset with clinician-written scenario pairs across eating, sexual safety, substance use and self harm, 12 frontier models.
.avif)