GPT-5.6 Sol solved a 30-year stats problem in 90 mins — by disproving a core conjecture
openai's gpt-5.6 sol pro disproved a 30-year-old statistics conjecture in 90 minutes, a task where gpt-5.5 failed after 20 hours, signaling ai's growing role in theoretical research.
researchers used openai's gpt-5.6 sol pro to construct a statistical model that disproves the assumption that the benjamini-hochberg procedure works reliably on correlated, continuous data, resolving a 30-year-old open question. the model achieved this in roughly 90 minutes, whereas gpt-5.5 failed to find a valid solution even after 20 hours of computation, marking a significant benchmark in ai capability. while the observed deviation in false discovery rates is relatively small, the finding proves that a widely used method can fail under specific conditions. experts note this signals a shift where ai can handle theoretical proof discovery, raising questions about the future role of human insight in statistical research.