GPT-5.6 Sol Pro Overturns 30-Year Statistical Conjecture in 90 Minutes

A statistics professor at the University of Pennsylvania used OpenAI's GPT-5.6 Sol Pro to disprove a central open conjecture about the Benjamini-Hochberg method in about 90 minutes, whereas the previous model GPT-5.5 failed to find an answer even after 20 hours. The solution reportedly combines known methods in a novel way, sparking public debate: can AI truly generate new knowledge, or merely recombine what it has learned? For researchers in math, statistics, or scientific discovery, this event suggests large models could become powerful tools for cracking long-standing problems. It also reminds AI developers to consider how to verify the reliability of "new knowledge" produced by models when they can overturn human-held conjectures in record time, and to adjust expectations of model capabilities accordingly.