Reasoning model

A reasoning model generates intermediate steps before its final answer, and is trained specifically to make those steps useful rather than decorative. On maths, multi-step logic and hard debugging this raises accuracy noticeably.

The trade is latency and cost, since those hidden steps are still generated tokens. For extraction, formatting, summarising and most writing, a standard model is faster and just as good.

These models also respond differently to prompting. Telling one to think step by step is redundant, and over-specifying the method can make it worse. Give the problem and the constraints, and let it work.