Asking a model to show its work isn't free, and it isn't a window into its mind — but on the right tasks, it measurably works.
A model answering directly jumps straight from a query to a final output token sequence, with no visible intermediate reasoning at all. Chain-of-thought (CoT) prompting instead asks the model to first generate an explicit step-by-step reasoning trace, and only then commit to a final answer. The difference isn't cosmetic. Every reasoning token the model produces on the way to an answer is additional computation it gets to condition on before it has to commit — which is exactly why CoT measurably improves accuracy on multi-step math, logic, and planning tasks. But that benefit isn't free, and it isn't what it looks like at a glance either: more tokens mean more latency and more cost per call, and a reasoning trace that reads as fluent, confident, and step-by-step is still not proof that the steps are logically valid or that the final answer is correct.