Chain of thought

Ask a model a hard question and it often does better if it works through the problem in writing first, one step at a time, before committing to an answer. That written working is called the chain of thought. Reasoning models do it by default and can spend a long time on it. Some products hide it, some show it, and some show a summarised version, which is why the same model can feel chatty in one app and terse in another.

The chain of thought is useful for more than accuracy. Because it is text, it can be read, and safety teams increasingly monitor it to spot a model heading somewhere it should not go and stop the run automatically. A caveat worth knowing: the written reasoning is not a guaranteed transcript of what the model actually did internally. It is a strong hint, not a confession.