GPT-5.6 Sol Pro has proven a result in mathematical optimization that has stood open since 1996, closing a 30-year gap in zeroth-order convex optimization [1]. The model produced the proof in a single 2.5-hour session from a 10-page prompt, and the argument was machine-verified in Lean [2]. The problem concerns deterministic zeroth-order convex optimization, and the result shows that d² function evaluations are necessary for this type of optimization [3]. The proof was produced by UC Berkeley IEOR teaching professor Phillip Kerger, who used a 10-page prompt to guide the model [4]. The result has significant implications for the field of mathematical optimization, and demonstrates the potential of AI-driven research in math and computer science [5].