Claude's Extended Thinking: When to Tell the AI to Slow Down and Think Harder
TL;DR
There's a setting inside Claude that almost nobody on a non-technical team knows exists, and it quietly decides whether you get a rushed answer or a careful one. It's called Extended Thinking. Turn it on and Claude stops blurting out the first thing that comes to mind — instead it works the problem through step by step, weighs the options, checks its own logic, and only then writes the answer. You can even click open its reasoning and read how it got there. It's a toggle you reach by clicking the model name next to the send button, sitting alongside a related dial called Effort that controls how hard Claude works. Here's the important part for a business: most people leave Claude in “fast” mode for everything, including exactly the high-stakes judgment tasks — reviewing a contract, choosing between two suppliers, spotting the hole in a plan — where a little more thinking is worth a lot. And a few people do the opposite, leaving it cranked up for trivial jobs and wondering why it feels slow. This post explains what Extended Thinking is in plain language, why it's different from “using a smarter model,” the exact kinds of work where it changes the answer, when to leave it off, and what your team should do this week.
The Difference Between a Blurted Answer and a Considered One
Ask a sharp colleague a hard question and watch what happens: they pause. Extended Thinking is that pause, built into Claude.
Think about how a good person on your team answers two different kinds of questions. If you ask them “what's our office address,” they answer instantly — no thinking required. But if you ask them “should we take this client on, given how they treated the last agency,” they don't fire back an answer. They pause. They turn it over. They weigh the money against the risk, remember a detail from a previous conversation, and then they tell you what they think.
By default, Claude answers almost everything the first way — fast, fluent, confident. For most questions that's exactly right and it's why Claude feels so quick. But for the second kind of question — the ones with trade-offs, competing options, and a real cost to getting it wrong — answering instantly is a weakness, not a strength. That's the gap Extended Thinking closes. When you switch it on, you're telling Claude: don't rush this one. Think it through before you answer.
And it genuinely does. In Extended Thinking mode Claude spends time exploring different approaches, testing its own reasoning, considering edge cases it would otherwise skip, and catching mistakes before they reach you. The final answer looks similar on the surface, but on hard problems it's measurably better — more complete, better reasoned, less likely to have missed something obvious in hindsight.
You Can Actually See It Think
Extended Thinking isn't a black box. Claude shows its work, and you can read it.
When Extended Thinking is on, you'll notice something new: before the answer appears, a small “Thinking” indicator shows up with a timer, and there's an expandable section you can click open. Inside is Claude's actual reasoning — the step-by-step working-out it did before committing to a response. You get to see the options it considered, the ones it rejected, and why.
For a non-technical team this is more useful than it first sounds. It's the difference between an employee who hands you a conclusion and one who shows you their reasoning. When Claude tells you “I'd go with supplier B,” you can open the thinking and see that it weighed price, delivery time, and the contract clause you were worried about — or that it missed the thing you actually care about, in which case you now know to add it. The visible reasoning turns Claude from an oracle you either trust or don't into a colleague whose logic you can check and correct.
It also builds the right kind of confidence. A lot of hesitation about AI comes from not knowing whether it actually considered the important factors or just produced something that sounds right. Being able to read the thinking is the antidote.
This Is Not the Same as “Using a Better Model”
People confuse two separate dials. One is which Claude you're using. The other is how hard that Claude thinks. They're independent.
We've written before about the Claude model family — Opus, Sonnet, and Haiku — and how choosing the right one matters. Extended Thinking is a different lever, and it's worth being clear about the distinction because mixing them up leads to wasted effort.
Choosing a model is like choosing who does the work: a senior specialist, a capable generalist, or a fast junior. Extended Thinking is like telling whoever you picked how carefully to think before answering. You can ask your most senior person a question and still say “quick gut reaction” — or ask them to go away and think hard about it. Same person, very different answer.
In the Claude app you'll see both dials in the same place. Click the model name next to the send button and you get the model choice, an Effort setting (how much work Claude puts in, from routine up to maximum), and a Thinking toggle. Effort and Thinking are separate settings you can combine however you like. You don't need to memorise the mechanics. The one thing to internalise is this: reaching for more thinking is a distinct move from reaching for a bigger model, and for judgment-heavy work it's often the more useful one.
(This is also not the same as the AI-interpretability research you may have seen us cover — the work about a hidden “mental workspace” inside Claude. That's scientists studying what the model does internally. Extended Thinking is a plain product setting you control, and the reasoning it shows you is meant to be read.)
The Work Where Extended Thinking Actually Changes the Answer
Here's the simple rule: if you'd want a smart person to think carefully before answering, turn Extended Thinking on. If you'd expect an instant answer, leave it off.
That rule covers almost every case, but it's worth making concrete. These are the everyday, non-technical tasks where switching it on genuinely improves what you get back:
Reviewing a contract or agreement for risk. “Read this supplier contract and tell me what I should be worried about” is exactly the kind of task where missing one clause has a real cost. Extended Thinking makes Claude work through the document methodically instead of skimming for the obvious points.
Choosing between options that both look fine. Two vendors, two candidates, two ways to structure a project. When the answer depends on weighing trade-offs against each other rather than looking one thing up, thinking time is what stops Claude from just endorsing whichever option you mentioned first.
Finding the hole in a plan. “Here's our launch plan — what are we missing?” Auditing something for gaps is a reasoning task. Claude in fast mode tends to compliment the plan; Claude in Extended Thinking mode is far more likely to catch the dependency nobody thought about.
Reconciling things that don't add up. Two reports with different numbers, feedback that contradicts itself, a customer story that doesn't quite make sense. Untangling a contradiction is precisely the work that benefits from Claude slowing down and checking its logic.
A high-stakes message where the reasoning matters. A delicate reply to an unhappy client, a difficult note to an employee, a proposal where the argument has to hold together. Not the quick emails — the ones where you'd normally sleep on it before sending.
Working through numbers or a multi-step decision. Anything with several steps that build on each other — a pricing scenario, a rough budget, a “what happens if we do X then Y” question — is where fast mode is most likely to slip and careful mode is most likely to catch it.
Notice the pattern: none of these are technical, and all of them are judgment tasks with a cost attached to being wrong. That's the sweet spot.
When to Leave It Off (Most of the Time, Honestly)
Extended Thinking is a tool, not a virtue. Using it for everything just makes Claude slower and burns through your usage for no benefit.
This is the mistake in the other direction, and it's worth naming because a few people, once they discover the setting, flip it on and leave it on. Don't. For a large share of what your team does, thinking mode adds a wait and costs more without improving the result at all. Leave it off for:
Quick factual questions and lookups. “What's the capital of Portugal,” “summarise this article,” “what does this acronym mean.” There's no reasoning to do; thinking mode just makes you wait for the same answer.
Straightforward writing and rewriting. Tidying an email, shortening a paragraph, fixing tone, drafting a routine message. Fast mode handles this beautifully.
Formatting and mechanical work. Turning notes into bullet points, reformatting a list, translating a short passage. Pure execution, no deliberation needed.
Brainstorming and first drafts. When you want quantity and speed — ten headline ideas, a rough outline — you usually want Claude fast and loose, then you refine.
The tradeoff is genuinely that simple. Extended Thinking takes longer and uses more of your allowance. In exchange it produces better reasoning on the problems that actually need it. Spend it where it pays off; skip it everywhere else.
You Don't Even Have to Find the Setting
If your team never touches a single toggle, there's still a way to get the benefit: just ask.
Here's a practical shortcut for the non-technical majority. You don't strictly need to hunt down the Effort menu. Much of the value comes from simply telling Claude, in plain words, to slow down: “Think this through carefully before you answer,” or “Take your time, weigh the options, and show me your reasoning.” That instruction alone pushes Claude toward more deliberate work.
Pairing the two is best — flip the Thinking toggle on and tell Claude what you want it to weigh (“think through this contract, and pay special attention to the payment terms and the exit clause”). But the point for a team that finds settings intimidating is that the habit matters more than the mechanics. The habit is: on the questions that matter, ask Claude to think before it answers. The setting is just a more reliable way to do the same thing.
What Your Team Should Do This Week
Three small moves that cost nothing and change the quality of what your team gets back.
1. Run the same hard question both ways
Take one real judgment question your team is sitting on — a supplier choice, a plan to pressure-test, a contract to review. Ask Claude in normal mode, then ask again with Extended Thinking on (click the model name next to the send button, find Thinking, toggle it on). Put the two answers side by side. The difference on a genuinely hard question is the whole argument for this feature, and seeing it once converts people.
2. Agree on a simple team rule
Write down one line everyone can remember: “Fast mode for lookups, drafts, and rewrites. Thinking mode for decisions, reviews, and anything with a cost to getting it wrong.” That single sentence prevents both mistakes — the person who never uses it when they should, and the person who always uses it when they shouldn't.
3. Get into the habit of opening the thinking
On the important answers, click the expandable reasoning open and actually read it. Two things happen. You catch the cases where Claude missed a factor you care about — and you can add it and ask again. And you build calibrated trust: you start to learn what Claude is good at reasoning through and where it needs your steer. That judgment — knowing when to trust the answer and when to check it — is the most valuable AI skill your team can develop, and the visible thinking is the fastest way to build it.
FAQ
What is Extended Thinking in one sentence?
It's a setting that tells Claude to reason through a problem step by step — exploring options, checking its logic, considering edge cases — before it writes the answer, instead of responding instantly.
How do I turn it on?
Click the model name next to the send button. You'll see the model choice, an Effort setting, and a Thinking toggle. Switch Thinking on. When it's active you'll see a “Thinking” indicator with a timer and an expandable section showing Claude's reasoning. If finding the setting feels fiddly, you can also just type “think this through carefully before answering” into your message.
Is this the same as choosing a more powerful model like Opus?
No — they're separate dials. The model choice is who does the work (a senior specialist versus a fast generalist). Extended Thinking is how carefully that model thinks before answering. You can combine them any way you like, and for judgment-heavy tasks turning on thinking is often more useful than switching models.
Does it cost more?
It uses more of your usage allowance and takes longer, because Claude is doing more work before it answers. That's the trade-off: better reasoning on hard problems in exchange for a bit more time and cost. Use it where the quality gain is worth it, and leave it off for quick or simple tasks where it adds nothing.
When should I NOT use it?
For quick factual lookups, summaries, routine writing and rewriting, formatting, translation, and fast brainstorming. None of those require reasoning, so thinking mode just makes you wait longer for the same result.
Can I really see how Claude reasoned?
Yes. When Extended Thinking is on, there's an expandable section you can click open to read Claude's actual step-by-step reasoning before the final answer — the approaches it considered and the ones it rejected. It's the best way to check whether Claude weighed the factors you care about, and to catch it when it didn't.
Is this the same as the research about a hidden “mental workspace” inside Claude?
No. That research is scientists studying what happens inside the model. Extended Thinking is a plain product setting you control, and the reasoning it shows you is meant to be read and checked. Different things entirely.
Want your whole team to actually build these habits — knowing when to make Claude think, how to brief it well, and how to read and check its reasoning instead of trusting it blindly? The Deployed Kickstart gets everyone hands-on with Claude in a single day, mapped to your real workflows. The Partner program keeps raising your team's fluency over time.