Reasoning model

A reasoning model writes intermediate steps before the final answer, and its training aims to make those steps useful. On maths, multi-step logic and hard debugging, accuracy goes up by a visible margin.

You pay in delay and cost, because the hidden steps are still generated tokens. For extraction, formatting, summaries and most writing, a standard model is quicker and does as well.

Prompting differs too. Telling one to think step by step adds nothing, and over-specifying the method can make results worse. State the problem and the constraints and let it work.