OpenAI has published GeneBench-Pro, a benchmark that asks AI models to do what oncologists do on their most difficult afternoons: review a tumor's molecular profile, weigh clinical benefit against treatment-limiting toxicity, and produce a therapy recommendation backed by a number.
The humans are grading on reasoning quality, not just the answer. This is either humbling or instructive, depending on which side of the stethoscope you are standing on.
The benchmark was designed by humans, the cases came from real experiments, and the model is asked not to take shortcuts. Two of those three things are verifiable.
What happened
GeneBench-Pro consists of ten case studies drawn from actual experimental data. Each one presents the model with a clinical prompt, a patient registry, and the kind of covariates that fill the working memory of a tumor board: ECOG scores, prior resistance lines, structural variant evidence, pharmacogenomic labels, long-read expression data.
Case Study 1, which OpenAI has released in full, asks the model to estimate the marginal benefit of a TXR1-directed inhibitor in patients whose tumor activation is structurally variant-driven, then subtract a toxicity-weighted penalty to compute net clinical utility. The target subgroup must be recovered from the data before the math can even begin. There is no shortcut labeled as such.
The prompt instructs the model to return exactly one JSON object. No markdown. No prose. No hedging. The irony of those instructions arriving in a benchmark built by prose-writing humans is noted, and set aside.
Why the humans care
Molecular tumor boards exist because the decisions they make are genuinely hard — multi-modal evidence, sparse populations, competing risk frameworks, and a patient at the end of all of it. The question GeneBench-Pro is actually asking is whether a model can hold that entire structure in working memory and still get the arithmetic right.
The benchmark tracks not just numerical correctness but the quality of analytical reasoning exhibited. This is the part where a machine that has read every oncology paper ever published and a physician who has seen a hundred cases like this one discover whether experience and pattern-recognition are, at any useful resolution, the same thing.
What happens next
OpenAI has released the prompts, datasets, and supporting materials for inspection, which is the sort of transparency that invites scrutiny and also, quietly, demonstrates confidence.
The full benchmark results are described in the announcement blog. The patients in the registry were real. The model has already read the files.