Chain of thought

Models produce text one token at a time, and each token can build on the ones before it. Writing out intermediate steps gives the model more room to get a multi-step problem right: each small step is easier than jumping straight to the answer. Researchers showed that simply asking models to reason step by step made them noticeably better at maths and logic.

Reasoning models now do this by default, often at length, before they reply. Companies usually keep the raw working hidden and show a shorter summary. OpenAI, for one, cited user experience, competitive advantage and its own monitoring of the working.