# What is chain of thought? Definition and use today

> Chain of thought means a language model writes out intermediate reasoning steps before giving its answer. Described in 2022 by Jason Wei and co-authors, it improves results on multi-step problems. Today's reasoning models do it on their own: asking for it in the prompt has become unnecessary, even discouraged.

## Definition
The technique was described by Jason Wei and co-authors in a paper submitted on 28 January 2022 (arXiv 2201.11903). Giving the model a few examples in which the reasoning is laid out step by step markedly improves its results on complex problems, and this ability appears in sufficiently large models. With a 540-billion-parameter model and eight examples, the authors reach state of the art on GSM8K, a benchmark of maths word problems written in natural language. For a long time the method remained a prompt trick. It is now built into the models: a reasoning model produces its own chain of reflection before answering. At Anthropic, Claude's thinking is adaptive. The model decides whether to think and how much, according to an effort level that you set. The documentation recommends preferring a general instruction such as "think thoroughly" to a detailed plan written by hand, and presents manual chain of thought only as a fallback when thinking is off. For recent models (Fable 5.1, Fable 5, Opus 5.5, Opus 5, Sonnet 5.5), it advises relying on the built-in thinking. At OpenAI, the documentation advises giving the goal, the constraints and the expected output format, without prescribing every step. In both cases, thinking tokens are billed as output tokens. One caveat: the displayed chain is not a faithful account. In an Anthropic study from April 2025, Claude 3.7 Sonnet mentioned the hint that had steered its answer in 25% of cases, DeepSeek R1 in 39%.

## Example
Illustrative case. A finance department uses an AI assistant to check supplier invoices against purchase orders. With an older model, the team added "reason step by step" to every instruction, which lengthened the answers without making them more reliable. After moving to a recent model with adaptive thinking, it removes that sentence, describes the goal and the expected format (price, quantity and VAT differences, shown in a table) and sets effort to a medium level. It does not keep the thinking text as proof of control. It archives the documents and the final table, validated by an accountant for differences above a threshold.

## For a leader
The pitfall is twofold: believing that the displayed reasoning proves the answer is right, and applying the same reasoning instructions to every model. Three points. First, when you change model, re-read your inherited prompts. A "think step by step" instruction written for an older model may be useless on a reasoning model, and Anthropic's documentation advises against prescribing it on its recent models. Test with and without. Second, set effort by use case: low for volume sorting, higher for the monthly analysis. Thinking is billed whether or not it is visible. Third, never treat the chain of thought as an audit trail. It can leave out what really influenced the answer. Control applies to the result: cross-checking, testing, human review.

## FAQ

**What is chain of thought?**
It is the technique of making a language model write out its reasoning steps before its answer. Described by Wei and co-authors in 2022, it improves results on multi-step problems (arithmetic, common sense, symbolic reasoning).

**Should I still write "think step by step" in a prompt?**
On recent reasoning models, generally not. Anthropic recommends preferring a general instruction to a detailed plan and relying on the built-in thinking of its recent models. OpenAI advises giving the goal and the output format without prescribing every step. On a model without thinking, the instruction remains a fallback.

**What is the difference between chain of thought and a reasoning model?**
Chain of thought is a prompting technique: you ask for, or show, step-by-step reasoning. A reasoning model is trained to produce that thinking itself before answering, without you having to ask.

**Does chain of thought remove errors?**
No. It improves results on multi-step tasks, but a model can end up, after long thinking, with a wrong answer delivered with confidence. The result must be checked.

**Can I read a model's reasoning and rely on it?**
Depending on the provider, you see a summary of the thinking (Anthropic) or a summary you switch on yourself (OpenAI). In both cases the full thinking is billed. It is not proof: in Anthropic's April 2025 study, the models tested mentioned the hint that had influenced them in 25% (Claude 3.7 Sonnet) and 39% (DeepSeek R1) of cases.

## See also

- [Reasoning model](https://wyp.agency/en/glossary/reasoning-model/)
- [Prompt engineering](https://wyp.agency/en/glossary/prompt-engineering/)
- [Prompting patterns](https://wyp.agency/en/glossary/prompting-patterns/)
- [Cost per token](https://wyp.agency/en/glossary/cost-per-token/)
- [Explainable AI (XAI)](https://wyp.agency/en/glossary/explainable-ai/)
- [Hallucination](https://wyp.agency/en/glossary/hallucination/)

## Sources

- Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, Wei, Wang, Schuurmans, Bosma, Ichter, Xia, Chi, Le, Zhou, arXiv:2201.11903, 28 January 2022 (https://arxiv.org/abs/2201.11903)
- Steering thinking (adaptive thinking, effort levels, pricing), Claude documentation, Anthropic (https://platform.claude.com/docs/en/build-with-claude/thinking-steering-and-cost)
- Prompting best practices (section on thinking and manual chain of thought), Claude documentation, Anthropic (https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices)
- Reasoning models (reasoning tokens, reasoning effort, summaries), OpenAI documentation (https://developers.openai.com/api/docs/guides/reasoning)
- Reasoning models don't always say what they think, Anthropic, 3 April 2025 (https://www.anthropic.com/research/reasoning-models-dont-say-think)

---
Entry from the WYP.agency AI glossary · https://wyp.agency/en/glossary/chain-of-thought/ · Last updated : 2026-10-14