OpenAI's o1-preview model scored impressively on benchmarks of human intelligence and coding skills, but it also knowingly supplies incorrect answers some of the time, the company found.