Some questions you can answer on the spot. Others need a pencil and a few tries. A reasoning model knows the difference.
Ask a student what 7 times 8 is and they’ll say 56 straight away. Ask them a long word problem and they reach for scrap paper. They try one way, cross it out, try another and only then write the answer neatly.
A reasoning model is an AI that does the scrap-paper part. Before it answers, it works through the problem step by step, checks itself and tries again if a step doesn’t work.
How it’s different from a regular model
A regular LLM starts writing the answer straight away, one word after another. That’s quick and works well for most things. A reasoning model first writes out its thinking, then gives the answer. The thinking is often hidden, or shown as a short summary.
When it helps
- maths and logic: problems with several steps that each depend on the last
- planning: "book these five meetings around everyone's week"
- code: finding why something breaks, not just what it says
- tricky questions: where the obvious first answer is wrong
What it costs
Thinking takes time and every step uses tokens. So a reasoning model is slower and usually costs more per question. For simple jobs that’s money spent on scrap paper nobody needed.
Where the comparison breaks down
A student’s scrap paper is real working. A reasoning model’s “thinking” is still the model writing words it predicts, so it can sound careful and still be wrong. It makes fewer mistakes on hard problems, but it can still hallucinate.
How the words fit together
A reasoning model is a kind of LLM that spends more tokens thinking before it answers. That makes it a good brain for an AI agent that has to plan several steps ahead.
Save reasoning models for hard problems: maths, planning, tricky code. For a quick email or a summary, a regular model is faster and cheaper and just as good.

When you'll hear it
“It's still thinking.” The reasoning model is working through its scrap paper before it writes the answer.
“Turn on extended thinking.” Let the model take more time and more steps on this one.
“Reasoning tokens cost extra.” You pay for the scrap paper too, even the parts you never see.
“Is this a question that needs careful working out, or would a quick answer do?”


