GPT-5.6 Sol reportedly disproves a 30-year-old statistics conjecture in 90 minutes after humans couldn't crack it

Editor's take

A recent report indicates that an advanced iteration of OpenAI's language model, GPT-5.6 Sol, successfully disproved a 30-year-old statistical conjecture concerning the Benjamini-Hochberg method. This achievement, accomplished in approximately 90 minutes, contrasts sharply with the inability of its predecessor, GPT-5.5, to find a solution even after extended computational effort.

This development signifies a potential leap in AI's capacity for abstract reasoning and problem-solving within highly specialized scientific domains. The Benjamini-Hochberg method is fundamental in multiple testing procedures across fields like genomics and clinical trials, meaning a formal resolution to this conjecture could have tangible impacts on scientific research rigor. The success here suggests AI models are moving beyond pattern recognition to genuine deductive capabilities.

Future observations should focus on whether this performance is reproducible across different complex mathematical problems and if similar capabilities emerge in competing models from Anthropic or Google. The broader question remains: what other long-standing scientific challenges might be susceptible to AI-driven breakthroughs, and what is the scalable pathway to identifying and posing these problems to AI systems?