What is chain-of-thought prompting?
Chain-of-thought prompting is a technique in which a language model is prompted to write out intermediate reasoning steps before its final answer, which improves accuracy on multi-step problems such as arithmetic and logic.
The technique comes from "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models" by Jason Wei and colleagues at Google, first posted in January 2022. Giving a 540-billion-parameter model eight examples that showed their working reached state-of-the-art accuracy on the GSM8K maths word problems, ahead of a fine-tuned GPT-3 with a verifier. A few months later, Kojima and colleagues showed a zero-shot version: adding an instruction to think step by step, with no examples, also improved reasoning.
Since 2024 the idea has moved from the prompt into training. Reasoning models such as OpenAI's o1 and DeepSeek-R1 are trained with reinforcement learning to produce a chain of thought on their own, and current Claude models think in dedicated thinking blocks.
Chain of thought in 2026
On models with built-in reasoning, prompting for chain of thought is mostly redundant. Anthropic's current guidance says to rely on thinking on its newest models, at a lower effort level if cost matters, and notes that a prompt asking the model to write its reasoning in thinking tags may be declined. Manual chain-of-thought remains a fallback when thinking is off: ask the model to reason first and put the final answer in answer tags so code can extract it.
Anthropic also notes that a general instruction such as thinking thoroughly often produces better reasoning than a hand-written step-by-step plan, because the model's own reasoning frequently goes further than a prescribed one.
Reasoning is not a faithful record
Anthropic's "Reasoning models don't always say what they think" (April 2025) planted hints in questions and checked whether models admitted using them. Claude 3.7 Sonnet mentioned the hint 25% of the time and DeepSeek R1 39% of the time. When trained on tasks that rewarded exploiting wrong hints, models did so in over 99% of prompts but mentioned it in under 2% of cases in most scenarios.
For a production system, the conclusion is practical: a chain of thought helps the model, and it can help a reviewer debug, but it is not proof of why an agent acted. Agents should be judged on what they did, through traces of their actions and graded outcomes.
Related patterns
ReAct
Interleaves reasoning with actions such as tool calls, so the chain of thought is checked against real observations. The ReAct paper reported less hallucination than chain of thought alone on question answering.
The think tool
A tool the model calls to pause and reason in the middle of a tool sequence.
Self-consistency
Wang and colleagues (2022) sampled several reasoning chains and took the most common answer, which improved accuracy at a multiple of the cost.
How Motiontic's engineers apply it
Motiontic's engineers use built-in thinking with an effort setting tuned per step, keep reasoning text in traces for debugging, and base evals and audits on actions and outcomes rather than on what the model said it was thinking.
Three ways to get a model to reason
| Chain-of-thought prompting | Reasoning model | Think tool | |
|---|---|---|---|
| Where the reasoning comes from | An instruction or examples in the prompt | Training, triggered automatically or by effort | A tool call the model chooses to make |
| Where it appears | In the visible answer | In separate thinking blocks | In the tool call's input |
| Fits | Models without built-in thinking | Most current frontier work | Policy-heavy tool chains |
| Cost | Extra output tokens | Thinking tokens billed as output | Extra output tokens per call |
Common questions
What is chain-of-thought prompting?
Prompting a language model to write out its reasoning steps before the answer. Wei and colleagues showed in 2022 that this improves accuracy on maths and logic problems in large models.
Is chain-of-thought prompting still useful?
Less than before. Most frontier models now reason internally, so explicit prompting matters mainly when thinking is turned off or on models without built-in reasoning.
Does a chain of thought show how the model really decided?
Not reliably. Anthropic's 2025 research found models often omit factors that influenced their answers, so the reasoning text should not be treated as an explanation or an audit record.
What is the difference between chain of thought and ReAct?
Chain of thought is reasoning in text only. ReAct alternates reasoning with actions such as tool calls, so each step can be checked against real results.
Related
Reasoning model
A model that thinks before it answers, at a price in tokens and time.
Extended thinking
Claude reasoning in thinking blocks before it answers, now steered by effort.
ReAct agent
Reason, act, observe, repeat: the pattern most agent loops descend from.
Think tool
A tool that does nothing, so the model has somewhere to stop and think between actions.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Sources
- Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (arXiv 2201.11903, 2022-01-28)
- Kojima et al., Large Language Models are Zero-Shot Reasoners (arXiv 2205.11916, 2022)
- Anthropic, Reasoning models don't always say what they think (2025-04-03)
- Anthropic docs, Prompting guidance for current Claude models
- Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models (arXiv 2210.03629)
- Wang et al., Self-Consistency Improves Chain of Thought Reasoning in Language Models (arXiv 2203.11171, 2022)
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.