When to switch on AI’s “thinking” mode
Thinking modes make an AI work through a problem before it answers. They help with maths, planning and checking, but mostly slow you down on simple jobs. Here’s how to choose.
Video transcript
Many AI assistants can now stop and think before they answer. It’s slower, but on the right problem it’s worth the wait. Here’s when to use it.
With thinking on, the model works through the problem before it replies: planning steps, trying approaches and checking its work. Some apps show a summary.
Use it for problems with several steps: maths and logic, plans with dates and budgets, comparing options, checking documents and code. For quick facts, rewrites or a chat, leave it off.
There are trade-offs. It’s slower. On some plans it uses up your allowance faster. It can still be wrong. And its answers can run long.
Prompt it differently, too. Don’t write out the steps for it. Give it the goal, the context and the limits, and let it work out how.
A good answer meets every constraint, flags the tricky days and lists its assumptions, so you know exactly what to check.
And check the answer anyway. Thinking cuts some mistakes, but not all. Test a step or two yourself, and ask it to show any sums.
Use thinking for the hard problems and leave it off for the rest. The full guide below has prompts to try.
In 30 seconds
- Switch thinking on for multi-step problems: maths, logic, plans with constraints, comparisons and checking documents.
- It’s slower, can use up your allowance faster, and can still be wrong.
- Give it the goal, context and constraints, not step-by-step instructions.
Many AI assistants can now stop and think before they answer. In some apps it’s a button or a menu option, in others a setting for how much effort to spend. On the right problem, the wait pays off. On the wrong one, you’re just waiting.
What thinking mode does
With Jargon busterThinking mode: A setting that makes an AI work through a problem before it answers. It’s slower, but often more accurate on maths, logic and planning. on, the model works through the problem before it replies: breaking it into steps, trying different approaches and checking its work. That working is extra text the model writes for itself, measured in Jargon busterToken: A chunk of text, often part of a word, that a language model reads and writes one at a time., so it costs time and computing power.
Some apps show a summary of the thinking, which you can open to see how it got there. Treat it as a rough guide, not proof. Anthropic, which makes Claude, has said it can’t be certain that a model’s visible thinking reflects what’s really going on inside it.
ChatGPT, Claude and Gemini all let you ask for more thinking, though the names and controls differ and change often. Some newer models also decide for themselves how much to think, depending on how hard the question looks.
When it helps, and when it doesn’t
| Switch it on | Leave it off |
|---|---|
| Problems with several steps, such as a household budget | Quick facts, such as a capital city |
| Maths, logic and puzzles | Rewording or tidying an email |
| Plans with constraints: dates, budgets, people | Brainstorming names or ideas |
| Comparing options against criteria | Casual chat or a quick translation |
| Checking a document for errors and gaps | Summarising a short article |
| Writing or fixing code | Simple formatting, such as a list or a table |
A quick test: would a person need a pen and paper, or a second look? If so, thinking is likely to help. If you’d answer off the top of your head, it won’t add much.
The trade-offs
- SlowerAnswers can take noticeably longer to arrive, especially on the higher settings. Fine for a plan, frustrating for a quick question.
- Uses more of your allowanceThinking uses extra tokens, so on some plans you’ll reach your Jargon busterUsage limit: A cap on how much you can use an AI tool, such as messages or uploads, in a set period. Once you reach it, you wait for it to reset. sooner.
- Still fallibleIt cuts some mistakes, not all. It can still state something false with confidence, a Jargon busterHallucination: When AI states something false as if it were true, because it predicts plausible words rather than checking facts., backed by careful-looking working.
- Longer answersReplies can be longer and more detailed than you need. Say what you want, such as “five bullet points”.
How to prompt a thinking model
You don’t need to tell it how to think. Give it the goal, the context and the constraints, plus what a good answer looks like, and let it work out the steps. Anthropic’s advice for Claude is that a general instruction to think carefully often beats a step-by-step plan you write yourself.
Goal: [a week of childcare cover in the school holidays]. Context: [two working parents; I’m at home on Mondays and Tuesdays; grandparents can do Wednesdays]. Constraints: [under £150 in total; no more than two days at holiday club; pick-ups by 5.30pm]. Give me a day-by-day plan in a table. Flag any day that doesn’t work, and list the assumptions you made.
Read the attached [document, such as a project plan or a contract]. My goal is to [spot anything that could cost me time or money]. List problems, gaps and contradictions, each with the page or section it’s on, most important first. Say which points you’re unsure about.
Then check its working as you would anyone’s: test a step or two yourself, and ask for the sum behind any number you’ll rely on.
Try it yourself 5 minutes
- Pick a fiddly everyday problem, such as a rota for four people with different days off.
- Ask it in a new chat with thinking off, or on its lowest setting, if your app allows.
- Ask again in another new chat with thinking on, using exactly the same words.
- Compare the two: which met every constraint, and how much longer did it take?
Check yourself
3 quick questions nothing is savedTools in this guide
Sources (4)
- Change the model, effort, and thinking settingsAnthropic (Claude Help Center)
- When should I use web search, extended thinking, and research?Anthropic (Claude Help Center)
- Claude’s extended thinkingAnthropic, February 2025
- Prompting best practicesAnthropic
Spotted a mistake? Tell us and an editor will check it.
