Introducing GeneBench-Pro

OpenAI has introduced GeneBench-Pro, a challenging research-level benchmark designed to evaluate AI performance in genomics, quantitative biology, and translational medicine. Building on the original GeneBench, the new benchmark assesses system-level judgment calls required in real-world computational biology, such as handling ambiguity, revising assumptions, and choosing correct analysis paths. GeneBench-Pro consists of 129 questions built using synthetic data with known causal structures, enabling deterministic grading against known targets and avoiding common benchmark failures associated with messy historical datasets. External domain experts reviewed problems for realism and appropriate methods, and OpenAI has open-sourced 10 representative questions on Hugging Face while providing a 50-question subset to Artificial Analysis. During testing, OpenAI's GPT-5.6 Sol model attained a pass rate of 28.7% at its highest reasoning level, increasing to 31.5% with Pro mode enabled. This represents a significant increase from earlier models like GPT-5, which scored below 5% on the original GeneBench. Results also highlight the impact of scaling test-time compute, with lower reasoning levels yielding single-digit pass rates. While current frontier models still solve fewer than a third of the problems, human experts estimate that a typical problem takes 20 to 40 hours to complete, suggesting that even partial automation at current capabilities could create meaningful economic and scientific value given inference costs of several dollars per problem.