You’ve seen the label on newer models: “reasoning.” It sounds like marketing, and some of it is. But there’s a real change underneath, and it’s worth understanding because it affects when you’d want to use one.
The short version
A standard model answers in one pass — it reads your question and produces a response straight through. A reasoning model is trained to work through a problem in steps first, almost like showing its work on scratch paper, before giving you the final answer. That extra thinking happens behind the scenes and takes a little longer.
When it helps
The step-by-step approach pays off on problems with a chain of logic: math, multi-part questions, debugging, anything where a small early mistake wrecks the final answer. On simple lookups or quick rewrites, it’s overkill — you wait longer for no real benefit.
The honest trade-off
Reasoning isn’t magic, and it isn’t always right. It reduces certain kinds of careless errors; it doesn’t give the model new facts it never had. Think of it as a model that slows down to double-check itself — useful when the question is hard, unnecessary when it isn’t.