Chain of thought
A prompting technique (or trained behavior) where the model produces intermediate reasoning steps before its final answer, which measurably improves accuracy on hard problems.
Chain-of-thought prompting was introduced as a research finding in 2022: for hard reasoning problems, asking the model to "think step by step" before answering produced meaningfully better answers than asking for the answer directly. The visible reasoning steps aren't decoration — they seem to be doing computational work the model needs to arrive at the right answer.
Frontier models like Claude Sonnet, GPT-5, and Gemini 3 have chain-of-thought behavior baked in. On hard problems they will produce reasoning steps unprompted; on easy ones they'll skip straight to the answer. Some products expose an explicit "thinking" mode that budgets tokens for extended chain-of-thought before generating the visible answer.
The active research question in 2026: how well does the visible chain of thought correspond to what the model actually computed? Recent interpretability work suggests the two can diverge — the visible reasoning is sometimes a post-hoc rationalization rather than the causal path to the answer. This has implications for AI safety and for how much you should trust chain-of-thought as an explanation.