What is a reasoning model?#
Ordinary models start writing an answer almost immediately. A reasoning model goes through intermediate steps before the final answer: it breaks the problem down, tries solutions, corrects its own mistakes and then answers. This "thinking time" raises accuracy on some problems.
An example#
Give a multi-step math problem or a bug hidden in complex code to a reasoning model. Instead of a quick guess, it analyzes step by step and usually has a better chance of giving the right answer.
Where it is useful#
- Math and logic
- Debugging and algorithm design
- Multi-step planning
- Analyzing texts with several conditions
Where it is not needed#
For simple rewriting, translation or a short answer, a reasoning model is usually slower and more expensive without a noticeable advantage.
Advantages and costs#
- Better accuracy on hard problems.
- Slower responses and more token usage.
- It can still be wrong; fluent reasoning is not necessarily correct.
Practical tips#
- Use a faster model for simple tasks.
- Write the problem fully and clearly; there is no need to say "think step by step", since it already does.
- Check the final answer, especially in calculations and code.
Relation to chain of thought#
The basic idea is chain of thought, except that these models are specially trained to think, so you do not need to ask for it in the prompt.