GeneBench-Pro
Research-level agentic evaluation suite from OpenAI measuring how AI agents perform multistage statistical reasoning over messy computational-biology and genomics data. 129 hard problems across 10 domains and 21 subdomains (statistical genetics, somatic oncology, single-cell and functional genomics, proteomics, clinical diagnostics, pharmacogenomics, population and structural genetics), with synthetically generated data of known causal structure graded deterministically via JSON.
Composite
71.4
Experimental validation
N/A — synthetic-data reasoning benchmark with known causal ground truth
Stages
Disease modelingTarget IDPhase II
Modalities
small molecule
Task types
classificationregressionretrieval
Size
problems: 129
public_problems: 10
held_out_third_party: 50
internal_holdout: 69
domains: 10
subdomains: 21
splits: {'train': 0, 'val': 0, 'test': 129}
note: 82 of 129 problems expert-reviewed; reviewers estimated 20-40h expert labor per problem; 10 public tasks released on HuggingFace under MIT
public_problems: 10
held_out_third_party: 50
internal_holdout: 69
domains: 10
subdomains: 21
splits: {'train': 0, 'val': 0, 'test': 129}
note: 82 of 129 problems expert-reviewed; reviewers estimated 20-40h expert labor per problem; 10 public tasks released on HuggingFace under MIT
License
MIT (public split) — remainder held out
First release
2026-06-30
Last updated
2026-06-30
Official site
Leaderboard
Dataset
Code / GitHub
→ repository
HuggingFace
Paper
GeneBench-Pro: A Research-Level Benchmark for AI Agents in Computational Genomics · OpenAI · 2026 · paper · doi:N/A — OpenAI technical report / bioRxiv 10.64898/2026.06.29.735386 · 0 citations
Flags
none
Experts
—
Groups
Hosted by
—
Related benchmarks
Rubric (7-criterion)
rigor
4
coverage
4
maintenance
3
adoption
2
quality
5
accessibility
3
industry_relevance
4
Notes
High-quality synthetic-with-known-ground-truth design enables deterministic, contamination-resistant grading (quality 5, rigor 4). Broad genomics/clinical coverage across 10 domains and expert-reviewed difficulty (coverage 4; reviewers estimate 20-40h/problem). Industry relevance 4: frontier-lab benchmark already used for cross-model comparison via Artificial Analysis, with top models still failing ~two-thirds of tasks (peak pass ~31.5%). Adoption 2 as a brand-new (Jun 30 2026) release; accessibility 3 since only 10 of 129 problems are public. Not self-referential in the pharma sense but authored by a model vendor — interpret cross-model rankings with that context.