According to Dynamic Insight Beating monitoring, OpenAI has released the Computational Biology Evaluation Benchmark GeneBench-Pro to test AI agents' multi-step decision-making ability in complex research scenarios such as genomics and translational medicine. The new benchmark consists of 129 questions (82 of which have been reviewed by external experts), and generates data with clear causal relationships through computer simulations to prevent models from taking shortcuts or catering to the preferences of the question setters to cheat.
The test results show that top models still struggle significantly when dealing with scientific reasoning involving quantifiable uncertainty. The strongest performer, GPT-5.6 Sol, only achieves a 31.5% accuracy rate in Pro mode, while Claude Opus 4.8 achieves a mere 16.0% accuracy. The research team points out that models commonly exhibit a disconnect of "being able to spot anomalies but not correcting the subsequent analysis," often choosing the wrong statistical methods or persisting in incorrect research directions.
