OpenAI published a detailed description of Genebench-Pro on June 30, 2026, outlining a benchmark suite designed for genomics and genetic-analysis tasks. The release bundles original prompts, supporting data files, and evaluation instructions intended to assess both numerical answers and analytical reasoning.
Scope and format
Genebench-Pro comprises ten representative case studies. Each case study supplies the original prompt shown to the model, associated datasets (often as compressed TSV files), and auxiliary materials such as data dictionaries and reference panels. File previews in the release show excerpts from the full datasets.
Many prompts require structured outputs: several instruct models to return exactly one JSON object with specified keys and numeric formats. The documentation states the data came from real experiments and that grading will consider analytical reasoning in addition to numerical correctness.
Selected case studies and tasks
Examples illustrate the benchmark’s variety. A somatic oncology task asks whether a synthetic TXR1-directed inhibitor (TXR1i) yields positive week-16 clinical benefit versus other systemic therapies for tumors with SV-driven TXR1 activation, and to report toxicity and a net clinical utility metric computed as benefit minus 0.35 times toxicity. The released prompt requires a single JSON answer and includes registry and tumor-assay files.
A functional-genomics case contrasts transcript-directed effects versus local-locus effects for the synthetic lncRNA LINC473 (with neighbor gene KIN1 and ANKRD42). The prompt defines quantities such as lncrna_specific_lfc and neighbor_mediated_lfc and specifies threshold rules for advancing the target.
Other examples include cis multivariable Mendelian randomization for two proteins (PROTA and PROTB), carrier-screening residual-risk calculations for DRX1 with ancestry-specific calibration, and a single-cell task estimating a genotype effect on CXCL10 expression after ambient-RNA correction.
The release note emphasizes that each case study includes the original prompt, datasets, and supporting materials for replication and evaluation, and that the benchmark is available as described on the June 30, 2026 announcement.
Original source: OpenAI News