GeneBench-Pro uses messy biological datasets to test whether models can make research decisions, not just answer textbook questions.
The strongest score in the first release was GPT-5.6 Sol Pro at 31.5% across 129 tasks; GPT-5.6 Sol reached 28.7%, while Claude Opus 4.8 led non-OpenAI systems at 16.0%. The benchmark stresses bias, failed quality control, statistical choices and the moment when an analysis plan must be abandoned.
Sources: CocoLoop、Introducing GeneBench-Pro (OpenAI); AlphaSignal; Silicon Report