Claude Extended Thinking: When to Make the AI Slow Down
Extended Thinking makes Claude work a problem through before answering — weighing options, checking its own logic — instead of writing the first thing that comes to mind. You reach it by clicking the model name next to the send button, alongside a related dial called Effort that controls how hard it works.
Most people leave it off for everything, including exactly the judgement-heavy tasks where it changes the answer: reviewing a contract, choosing between suppliers, finding the hole in a plan. In my experience it shifts the quality of a hard answer more reliably than switching to a more powerful model does — and almost nobody on a non-technical team knows it exists.
The pause a good colleague takes
Watch how a sharp person on your team answers two different questions. Ask “what is our office address” and they answer instantly. Ask “should we take this client on, given how they treated their last agency” and something different happens: they pause. They turn it over. They weigh the money against the risk, remember something from a previous conversation, and then tell you what they think.
Extended Thinking is that pause, made available as a setting. Without it, Claude answers both questions the same way — fluently and immediately. The fluency is the problem: a rushed answer to a hard question does not look rushed. It looks like an answer.
Why it is not the same as a smarter model
People conflate these constantly, and they are genuinely different levers.
| Choosing a more powerful model | Turning up thinking | |
|---|---|---|
| What it changes | How capable the reasoning can be | How much reasoning actually happens before answering |
| Costs you | More per answer, and some speed | Time — you wait longer for the same model |
| Best for | Problems that are genuinely difficult | Problems where the first instinct is likely to be wrong |
| Common mistake | Defaulting to the biggest model for everything | Never touching it at all |
The two compose. A capable model that answers instantly can still miss the thing you needed it to catch — not because it could not have found it, but because it did not look. That is why turning up thinking on the model you already have is worth trying before concluding you need to upgrade anything.
When it changes the answer
The pattern is consistent: it earns its time whenever the first plausible answer and the correct answer are different things.
- Anything adversarial. Reviewing a contract for what could hurt you, pressure-testing a plan, finding the flaw in an argument. Fast mode tends to summarise what is there. Thinking looks for what is missing.
- Decisions with competing constraints. Choosing between two suppliers where each wins on different dimensions. Weighing a hire. Anything where the answer is a trade-off rather than a fact.
- Multi-step reasoning. Working out a consequence three moves ahead, or anything where an early wrong turn quietly invalidates everything after it.
- Work where you cannot easily check the output. If you would not catch a subtle error, you want the version that checked itself.
Before you send, ask: would I be able to tell if this answer were subtly wrong? If the answer is no, turn thinking up. That single question routes almost every real decision correctly, and it is easy enough to remember that people actually use it.
When to leave it off
Being even-handed here matters, because the opposite mistake is real and it makes people quietly abandon the setting.
Leave it off for anything where the work is production rather than judgement: reformatting, summarising something you will read anyway, drafting a routine email, tidying notes, translating. Nothing improves by deliberating over how to lay out a meeting summary, and waiting for it is how people conclude the feature is not worth the bother.
Also leave it off when you are iterating quickly. If you are going back and forth refining something in ten short exchanges, the pause on every turn costs more than it returns. Turn it on for the pass that matters, not for the whole conversation.
Reading the reasoning
You can open up the reasoning and read how Claude got there, and this is underrated — not because you need to audit the logic, but for two more practical reasons.
It shows you what it misunderstood. When an answer is off, the reasoning usually reveals that it took your question to mean something slightly different. That is a briefing problem, and you can only see it by looking.
It shows you what it did not consider. If you were hoping it would weigh something and the reasoning never mentions it, you have learned something more useful than the answer itself: it did not have the context. That is your cue to supply it rather than to try the question again in different words.
What to do this week
- Find the setting. Click the model name next to the send button. Most people have never clicked there, and everything in this article lives behind it.
- Re-run one decision you already made. Take a judgement call from the last month where you used Claude and were unimpressed. Run it again with thinking turned up and compare. This is what recalibrates a team's sense of what the tool can do.
- Teach the one-sentence test, not the feature. Would I be able to tell if this were subtly wrong? A rule people remember beats a setting they were shown once.
Building the habit of knowing when to make Claude think — and how to read its reasoning rather than trusting the output — is the kind of judgement a Deployed Kickstart half-day builds against your team's real decisions. The Partner programme keeps raising that judgement over time.
Frequently asked questions
What is Extended Thinking in Claude?
A setting that makes Claude work a problem through before answering — weighing options and checking its own logic — instead of writing the first thing that comes to mind. It is the pause a sharp colleague takes before answering a hard question, made available as a toggle. You reach it by clicking the model name next to the send button, alongside a related dial called Effort.
Is Extended Thinking the same as using a more powerful model?
No, they are different levers. Choosing a more powerful model changes how capable the reasoning can be; turning up thinking changes how much reasoning actually happens before answering. A capable model answering instantly can still miss something — not because it could not have found it, but because it did not look. Try turning up thinking on the model you already have before concluding you need to upgrade.
When should I turn on Extended Thinking?
Whenever the first plausible answer and the correct answer are likely to differ: anything adversarial such as reviewing a contract for what could hurt you or pressure-testing a plan, decisions with competing constraints, multi-step reasoning where an early wrong turn invalidates everything after it, and work where you would not catch a subtle error yourself.
When should I leave Extended Thinking off?
For production rather than judgement — reformatting, summarising something you will read anyway, drafting routine emails, tidying notes, translating. Nothing improves by deliberating over how to lay out a meeting summary. Also leave it off while iterating quickly: if you are refining something across ten short exchanges, the pause on every turn costs more than it returns.
How do I know whether a task needs Extended Thinking?
Ask yourself one question before sending: would I be able to tell if this answer were subtly wrong? If the answer is no, turn thinking up. That test routes almost every real decision correctly and is simple enough that people actually remember to use it.
Why should I read Claude's reasoning?
For two practical reasons rather than to audit the logic. It shows you what Claude misunderstood — when an answer is off, the reasoning usually reveals it took your question to mean something slightly different, which is a briefing problem you can only see by looking. And it shows what it never considered, which tells you it lacked context rather than that it reasoned badly.
Found this useful? Send it to someone who needs it.