A reasoning model is a language model trained to write intermediate steps before giving its final answer. Those steps help it solve multi-step problems with fewer mistakes. In exchange, the answer takes longer and uses more tokens than one from a model that answers directly.
A question from the app. Pick an answer.
A reasoning model is like someone who...
That is answering with no draft. A reasoning model writes intermediate steps before the final answer.
It writes intermediate steps, then the answer. It is an analogy: it does not think like a person.
It does not look up a stored answer: it generates text, and when it reasons it also generates intermediate steps.
One way to picture it: someone who works things out on scratch paper before answering, instead of blurting out the first thing that comes to mind. It is only an analogy. The model does not think like a person: it is still a language model generating text, except that it first generates intermediate steps and then the answer.
It tends to help when the answer depends on chaining steps: logic problems, math, hard code, plans with constraints or decisions with several conditions. For simple questions, a known fact or rewording a text, a regular model often answers just as well, faster and cheaper. The safe way is to test on your cases.
Writing the draft takes time, because the answer is generated token by token. And those reasoning tokens are billed as output even when they are not shown. Many of these models offer an effort setting to choose how much to reason: less effort saves time and money, more effort can help on hard problems.
There is no guarantee of being right either: a reasoning model can be wrong, and a long draft does not make what it says true. Each provider names and configures these models in its own way, so check how the one you plan to use works.