CoT

The key finding was simple: if you ask a model to show its work instead of jumping straight to an answer, it makes far fewer mistakes on hard problems. In practice this means prompts like 'think step by step' or 'walk me through your reasoning.' The model generates intermediate reasoning steps that catch its own errors before reaching a conclusion.

In recent models, CoT has gone from a prompting trick to a core architectural feature. Models like o3 and Claude's extended thinking mode run an internal reasoning trace before outputting a response. You don't have to prompt them for it: they do it automatically on hard tasks, and you can sometimes see the reasoning in a separate 'thinking' pane. This is one reason reasoning models outperform standard models on math, coding, and multi-step logic.

For builders: CoT matters both as a prompting technique you can invoke and as a feature to look for in model selection. If your task involves complex reasoning, multi-step logic, or high-stakes decisions where errors are costly, preferring a model with built-in CoT or explicitly prompting for it usually improves output quality. The tradeoff is latency and tokens: thinking takes time and costs more.