OpenAI's GPT-5 Pro has independently derived a novel proof in convex optimization, improving a recent arXiv bound from 1/L to 1.5/L.