We provide the first demonstration that a neural network solvesuniversity-level mathematics problems. Our methods combinetwo innovations: (i) recent neural networks pre-trained on textand fine-tuned on code, rather than pre-trained on text alone,and (ii) novel techniques to automatically rephrase problems soneural networks can synthesize correct executable programs.We generate programs that perfectly solve a random sampleof problems from MIT mathematics courses including Singleand Multi-variable Calculus, Differential Equations, Probabilityand Statistics, Linear Algebra, and Mathematics for ComputerScience, as well as problems in the MATH benchmark of highschool math topics. Our methods also generate new questionsthat are indistinguishable by students from course questions.Implications for higher education include new roles of AI inautomatic course evaluation and content generation.