Chain-of-thought prompting asks a model to lay out intermediate steps before committing to an answer. The original trick was as blunt as appending "think step by step," and it worked: on multi-step arithmetic, logic, and analysis, models that reason aloud are measurably more accurate than models that answer immediately.
The technique has since been absorbed into the models themselves. A reasoning model does this internally by default, spending hidden tokens on deliberation before producing visible output — which is why those models cost more per answer and take longer to respond.
Why it matters at your desk. Two practical consequences. First, on anything with steps — a calculation, a comparison across documents, a decision with conditions — asking for the working usually improves the answer and always makes it checkable. A conclusion you can audit is worth more than a confident one you cannot.
Second, it explains the cost and latency profile of reasoning tiers. You are paying for thinking you mostly do not see. For straightforward retrieval or formatting, that spend buys nothing.
The caveat is sharper than it first appears: the stated reasoning is not a transcript of the computation. A model can produce plausible steps and a wrong answer, or a right answer via steps that do not actually support it. Visible reasoning is evidence to check, not proof. This gap is also why labs treat reasoning traces as a safety surface in their own right, and why safety reviews of reasoning-heavy releases have become a gating factor on shipping.
What to watch for: read the steps for errors rather than reassurance. Fluent reasoning is easy to generate; correct reasoning is the thing you are checking for.