The company's experimental reasoning model wasn't fine-tuned to solve math problems, but was trained as a general problem-solver.