Thinking before answering
A reasoning model is an LLM that takes time to work through a problem in stages before producing an answer. Vendors began shipping them around 2024, and they now sit alongside immediate-answer models as a second pillar. Thinking modes in ChatGPT and equivalent features elsewhere are examples.
How it differs
A conventional LLM starts generating an answer as soon as it receives the question. A reasoning model first works internally through how the problem decomposes and whether its approach is right, corrects itself when it notices an error, and only then produces the final answer.
It is the difference between answering immediately and working through the steps on paper. Answers take longer and cost more as a result.
Where it helps and where it does not
- Suits: mathematics, debugging, logic puzzles, complex planning, close analysis of long documents
- Does not suit: greetings, simple questions, casual conversation, anything where speed matters — the thinking time is wasted
Most services let you switch between immediate and reasoning modes, or choose automatically based on the question.
Using them well
Select a reasoning model only when you genuinely want the problem thought through. Everyday research is fine on a standard model, while complex analysis and difficult judgment calls are where reasoning models show their value. If an answer feels shallow, switching models is worth trying — it often changes the result substantially. Note that this is a different sense from inference, the general term for a model producing output.
The arrival of reasoning models also showed that capability can be increased by spending time at answer generation, not only by making models larger. That direction leads on to long-running research features such as Deep Research.