OpenAI releases GeneBench-Pro biology benchmark

Measuring expert-level workflow execution rather than static Q&A is the next frontier of LLM evaluation, and GeneBench-Pro shows how far we still have to go.

• –By using synthetic problems derived from known causal structures, the benchmark allows for deterministic grading of highly complex, open-ended tasks.
• –The stark contrast between GPT-5.6 Sol (31.5%) and prior models (under 5%) suggests that advanced reasoning architectures are starting to grasp multi-stage scientific workflows.
• –Forcing models to navigate noisy data and inferential forks exposes the limits of raw next-token prediction, emphasizing the need for robust planning and agentic execution.