OpenAI tested OpenAI's internal frontier model, which produced 722 AI-generated math proofs, across thousands of mathematical problems, spending the equivalent of approximately three hours of ChatGPT Pro reasoning on the average result