Chain-of-Thought
Getting a model to reason step by step before answering — which dramatically improves its performance on hard problems.
Reviewed July 10, 2026Stable
When not to use it
- On simple tasks. It adds tokens, latency, and cost for no gain — and on easy questions it can talk the model out of a correct first instinct.
- As an explanation of the model's actual process. The stated reasoning is generated text, not a transcript of computation. It can be plausible and unrelated to how the answer was reached.
- When you need short answers. Reasoning that leaks into the output is a formatting bug for most product surfaces.
Reach for something else instead
- Few-shot examples often get the same lift with fewer tokens.
- Tools — for arithmetic or lookups, let the model call a calculator or a database rather than reason through it.
- Decomposition in your code — separate prompts per step gives you control, checkpoints, and debuggability.
Read more on the blog
- Prompt engineering in 2026: what still works and what changedIf your prompts still open with "act as an expert" and "let's think step by step," you're using 2023 advice on 2026 models, and some of it now makes your output worse. What actually moves results today, what quietly stopped working, and why.
- How reasoning models work: AI that thinks before answeringIn 2025 a new kind of model arrived: one that pauses to think, working through a problem on internal scratch paper before answering. Reasoning models like o1, o3, and DeepSeek-R1 trade speed for accuracy on hard problems. Here is what they actually do, how they learned to do it, and when the extra thinking is worth it.
- How AI models are trained: from raw text to a system that helpsA language model isn't programmed, it's grown, in stages, from a firehose of text into a system that answers helpfully. The full modern pipeline: pretraining, supervised fine-tuning, preference alignment, and the reasoning training that defines 2026, what each stage does, why none can be skipped, and how the recipe changed.
- What is reinforcement learning? Learning from rewardReinforcement learning went from a niche corner of AI obsessed with games and robots to the paradigm that shapes how every modern language model behaves. Here's what it actually is, learning by trial, reward, and consequence, why it's different from other machine learning, and how it quietly became the layer between a smart model and a useful one.
Further reading
- Wei et al. (2022), Chain-of-Thought Prompting Elicits Reasoning in Large Language Models — the original.
- Kojima et al. (2022), Large Language Models are Zero-Shot Reasoners — the step-by-step result.
- Turpin et al. (2023), Language Models Don't Always Say What They Think — stated reasoning can be plausible and unfaithful. Read this before trusting a chain.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Trusting the reasoning because it sounds rigorous. Faithfulness of stated reasoning is an open research problem, not a solved one.
- Using it everywhere by default. Measure it; on many tasks it costs more and helps nothing.
- Showing the chain to end users, who reasonably read it as the system's real thinking.
At a glance
FieldLanguage & LLMs
Core ideareason step by step before answering
Benefitbig accuracy gains on hard tasks
Caveatreasoning may not be faithful
DifficultyBeginner → Intermediate
Where this sits
26 concepts come first. Understanding it opens up 7 more.
13Levelsteps in
26Needs firstconcepts
7Opens up2% of 310
3Areascrosses fields
Learn these firstIn-Context Learning
Computed from the prerequisite graph, not assigned. How this works