OpenAI Launches GeneBench-Pro to Test AI in Biology

OpenAI has unveiled GeneBench-Pro, a cutting-edge benchmark targeting AI performance in computational biology, genomics, and translational medicine. Released on June 30, 2026, this research-level tool is designed to push AI beyond routine data handling into the realm of judgment-heavy, real-world scientific reasoning.

GeneBench-Pro builds on the earlier GeneBench framework, now tackling more complex tasks across 129 meticulously designed problems. These problems simulate real-world scenarios, requiring AI systems to analyze messy datasets, revise assumptions, and make iterative decisions—tasks that mirror the nuanced workflows of human researchers. According to OpenAI, this benchmark aims to evaluate "research taste," or the ability to make informed judgment calls in ambiguous scientific contexts.

Unlike traditional benchmarks, GeneBench-Pro avoids pitfalls like arbitrary grading by synthetically constructing datasets with known causal structures. This allows deterministic grading and ensures that solutions depend on robust analytical reasoning rather than shortcuts or arbitrary choices. Problems span ten domains and 21 sub-domains, ranging from genomics to translational medicine, reflecting the breadth of computational biology challenges.

AI Performance and Challenges

Preliminary results highlight both progress and limitations. OpenAI's GPT-5.6 Sol model achieved a 28.7% pass rate at the highest reasoning level, improving significantly from earlier versions like GPT-5, which scored below 5% on similar tasks. With Pro mode enabled, the pass rate climbs to 31.5%. These numbers underscore the rapid evolution of frontier models, though the benchmark’s inherent difficulty leaves ample room for improvement.

Human reviewers estimate that solving a typical GeneBench-Pro problem would take an expert 20–40 hours, with labor costs exceeding $4,000 per problem. By contrast, AI inference costs are just a few dollars per problem, suggesting even partial automation could yield significant economic and scientific value. However, current models still struggle with "closing the inferential loop," a failure pattern akin to novice researchers who can identify patterns but lack the experience to contextualize findings effectively.

Validation and Future Applications

GeneBench-Pro problems underwent rigorous validation, including external reviews by domain experts such as graduate students and professors in fields like human genetics. Reviewers praised the benchmark’s realism, noting its alignment with the messy, iterative nature of actual biological research. Alexander Strudwick Young, an assistant professor at UCLA, remarked that the problems demand “thoughtful and reflective data analysis” beyond applying standard methods to clean datasets.

OpenAI has open-sourced 10 representative problems on Hugging Face and plans to release a 50-question subset for independent benchmarking via Artificial Analysis. These steps aim to foster collaboration and transparency while accelerating advancements in AI-driven scientific research.

Broader Implications

With sequencing costs plummeting and vast biobank-scale datasets becoming increasingly accessible, the bottleneck in biology has shifted from data generation to analysis. Accurate and efficient AI systems could transform this landscape by automating hypothesis testing, experimental design, and data interpretation. As Jennifer Grundman, a PhD candidate in human genetics at UCLA, noted, models that excel at GeneBench-Pro tasks could “greatly improve the pace, thoroughness, and reproducibility of research.”

GeneBench-Pro represents a significant step toward building AI systems capable of assisting—or even leading—complex scientific workflows. While current systems are far from replacing human expertise, their ability to handle judgment-heavy tasks offers a glimpse into a future where AI accelerates discoveries in medicine, biology, and beyond.