OpenAI Does Not Trust Its Own Models, and Rightly So

that's a new benchmark