Glossary
Chain-of-thought
Chain-of-thought is the intermediate working a model produces before its final answer, whether because you asked it to show its steps or because it was trained to produce them by default.
In plain terms
It looks like showing your working in a maths exam, and it is not quite that. In an exam the working is a record of thinking you already did. Here, writing the steps is what makes the better answer more likely in the first place, because each step becomes part of what the next one is built from. That is why asking for the working changes the answer rather than merely explaining it.
Why it matters
It is the cheapest quality improvement available to you. Adding a request to work through something step by step costs nothing, takes seconds, and makes a real difference on any task with more than one stage to it. It is also the best diagnostic you have: when an answer is wrong, the steps usually show you exactly where it turned, which beats re-reading the answer and guessing. And it is the thing your tools now label as thinking, so knowing what it actually is stops you reading more into that panel than it can support.
How it works
The steps are produced the same way as the answer, one piece of text after another, with nothing separate happening behind them. That is the whole mechanism, and the consequence is the interesting part: once a step has been written it forms part of what the model is continuing from, so a correct intermediate result makes the next correct result more likely. Writing 660 makes the sentence after it easier to get right.
It follows that the working is not a description of reasoning that occurred somewhere else. There is no hidden deliberation being summarised for your benefit. Insofar as there is deliberation at all, the text is it, which is why an answer produced with working can differ from the same question answered directly.
You can ask for it, or buy a model that does it unprompted. Asking any capable model to work through a problem in steps usually produces this behaviour and is free. A model trained for it produces the working by default and, more importantly, was trained towards working that arrives at correct answers rather than working that merely reads well.
It is output, so it is charged and it takes time. Where a product hides the steps, you are still paying for every one of them, and the wait you notice is largely them being generated. This is the usual reason a considered answer costs several times a direct one despite the visible reply being no longer.
Treat it as a signal rather than an explanation. It is genuinely useful for finding where an answer went astray, and researchers have found cases where the stated steps do not match what actually determined the output. A chain can be a plausible story told alongside the answer rather than the route to it, so it will support a diagnosis and not an audit.
One chain, and what each part of it is good for
Seen in the wild
Ask a question with two or three dependent steps, then ask the identical question again with a request to work through it in steps, and compare the two answers rather than the two explanations.
ChatGPTWhen a long answer comes back wrong, read the working above it and find the step where it turned, which is usually quicker than re-reading the answer.
ClaudeRun a model on your own machine and watch the steps arrive one after another, which makes it plain that they are being written rather than reported.
LM Studio
Common misconceptions
People assume
It shows me how the model thinks.
In fact
It shows text the model produced before its answer. Researchers have found cases where the stated steps do not match what actually drove the output, so a chain can be a plausible account rather than the route taken. Useful for spotting a wrong turn, unreliable as an explanation of the system.
People assume
Asking for the working just makes the reply longer.
In fact
It changes the reply. On anything with dependent stages, the accuracy moves, because each written step forms part of what the next is generated from. This is the rare case where a change to how you phrase a request alters whether the answer is right rather than how it reads.
People assume
If the product hides the thinking, I am not paying for it.
In fact
You are paying for all of it. Hidden working is generated exactly like visible working and billed the same way, and it accounts for most of the extra wait as well. Hiding it is a decision about the interface, not about the invoice.
Telling them apart
Chain-of-thought vs Reasoning model
Chain-of-thought
The working itself. Free to ask for, available from most capable models, and a prompting decision rather than a purchasing one.
A model trained to produce that working by default, and trained towards working that reaches correct answers rather than working that reads well.
If the fix is a sentence added to your request, you want the first. If step-by-step prompting has already been tried and the answers are still wrong on genuinely hard problems, that is the case for the second.
Questions
- Does asking for step-by-step working really help?
- On tasks with dependent stages, yes, and noticeably. On tasks with none, such as rewriting a paragraph or fixing tone, it adds length and delay and changes nothing worth having. The useful habit is reaching for it when a wrong answer would come from a missed intermediate stage rather than from missing knowledge.
- Why do some products hide the working?
- Partly because a raw chain can be long and unpolished to read, and partly because providers treat it as their own material. What is displayed is often a summary prepared for readability, which is worth knowing before you quote one in a report.
- Can I use the steps as an audit trail?
- No, and this is the boundary worth holding. The chain is not guaranteed to correspond to what determined the answer, so it will not support a claim about how a decision was reached. If you need defensible reasoning, the answer has to be checked against sources rather than against its own account of itself.
- Does it make hallucination less likely?
- On problems with checkable structure, generally yes, because a wrong intermediate result is easier to notice than a wrong conclusion. It does not help where the underlying issue is missing knowledge, and it introduces a failure of its own: a wrong first step gets elaborated carefully, which reads more convincingly than a quick mistake.
Key takeaways
- The steps are generated like any other text, and writing them is what makes the better answer more likely.
- They are not a report on thinking that happened elsewhere; there is no hidden process being summarised.
- You can elicit them free from most models, or buy a model trained to produce them by default.
- They are billed and timed as output whether or not the product shows them to you.
- Use them to find where an answer went wrong; do not use them as an audit trail.
Last checked July 2026