---
title: "Extended thinking: when deeper reasoning pays off"
description: "What \"thinking\" means in Claude Recent Claude models can reason before they respond. In the Claude apps this is the extended thinking option; in the API…"
url: https://optimizeall.com/learn/mastering-claude/extended-thinking-when-and-how
updated: 2026-10-05
---

Mastering Claude (Anthropic) · Deeper thinking and writing workflows · lesson 11 of 20 · 13 min

# Extended thinking: when deeper reasoning pays off

## What "thinking" means in Claude

Recent Claude models can reason before they respond. In the Claude apps this is the **extended thinking** option; in the API it is **adaptive thinking**, where the model decides how much to think, steered by an **effort** level (low, medium, high, extra high and max on the newest models). Newer models such as Claude Opus 5.5 think by default and let effort control depth; older integrations used a fixed "thinking budget", which Anthropic has deprecated on current models.

The raw internal reasoning is not shown in full. Depending on the product you may see a summary of the reasoning, progress notes during long tasks, or nothing. Summaries are useful for spotting wrong assumptions, but they are not a complete audit trail.

## When thinking pays off

| Worth it | Not worth it |
|---|---|
| Multi-constraint planning (budgets, schedules, staffing) | Rewriting a caption |
| Maths, pricing, forecasting logic | Translating a short message |
| Comparing options against weighted criteria | Formatting a list |
| Debugging a formula chain or code | Simple lookups |
| Multi-document analysis with contradictions | Brainstorming 20 hashtags |
| Long agentic tasks (coding, research) | Quick factual Q&A with search |

The trade-off: better answers on hard problems in exchange for more time and more usage. Effort matters more on newer models than older ones; a lower setting on a new model often matches a higher setting on the previous generation.

## How to prompt for good reasoning

Thinking amplifies the quality of the problem statement. Give it:

1. **The goal and decision** ("Choose one of three retainer models for Q1").
2. **All the data** (costs, volumes, constraints), not a summary of it.
3. **Constraints** (cash-flow minimum, headcount cap, client contract terms).
4. **Success criteria** ("maximise margin while keeping churn risk low; explain trade-offs").
5. **Output shape** (a comparison table, the recommendation, assumptions list, sensitivity check).

You do not need to tell current models *how* to think step by step; over-prescriptive instructions can make results worse. Tell them *what* a good answer must contain.

```text
<goal>Recommend one of three pricing models for our 6-person content agency for Q1.</goal>
<data>[paste costs, client list with monthly fees, hours per client, churn history]</data>
<constraints>
- Minimum monthly cash buffer of [amount]
- No new hires before March
- Two clients have fixed-price contracts until June
</constraints>
<criteria>Maximise gross margin; keep risk of losing top-3 clients low; simple to explain.</criteria>
<output>
1. Table comparing the three models on margin, risk and admin effort
2. Recommendation with reasoning
3. Every assumption you made, numbered, so I can challenge them
4. What would change your recommendation (sensitivity)
</output>
```

## Reading and challenging the result

- **Scan the assumptions list first.** Wrong assumptions produce precise-looking wrong answers.
- **Check the arithmetic** on two or three key figures (or ask Claude to recompute with code execution).
- **Correct and rerun:** "Assumption 4 is wrong: we work 22 days a month, not 20. Rerun the comparison."
- **Ask for the counter-case:** "Argue for the option you did not choose."

## Worked example: freelancer capacity planning

A video editor in Lahore with clients in the UAE and UK asks whether she can take on a retainer worth four edits a week. With thinking on and full data (current commitments, average edit hours, revision rounds, public holidays in both countries), Claude builds a capacity table and flags that two UK clients' revision rounds make one week in December infeasible. The assumptions list shows it treated a Pakistani public holiday as a working day; she corrects it, reruns, and negotiates a revised start date with the new client.

## Hands-on

1. Pick one real decision you face this month.
2. Run the template with extended thinking **off**, then **on** (same model, fresh chats).
3. Compare: which assumptions differ? Which answer would you act on? Was the extra time worth it?
4. Record the result in your model-comparison table from Lesson 1.

## Pitfalls

- Using maximum thinking for everything and hitting usage limits by lunchtime.
- Missing context: thinking cannot invent your cost data.
- Treating a reasoning summary as proof. It shows direction, not verification.

## How to measure success

For decisions you run with thinking, the assumptions list catches at least one issue you would have missed, you verify key figures before acting, and you reserve thinking for tasks where it demonstrably changes the answer.

## Video lecture: Extended thinking: when deeper reasoning pays off

Lecture coming soon · 14 chapters · about 9 minutes. Read the full transcript below.

1. Thinking before answering
2. Why thinking changes answers
3. What thinking is
4. Worth it vs not
5. Thinking amplifies the problem statement
6. Tell it what, not how
7. Challenge the result
8. Simple example
9. Worked example: capacity planning in Lahore
10. Pitfalls
11. Try this now
12. Watch me do it, part 1
13. Watch me do it, part 2
14. Recap and next step

## Lecture transcript

### Thinking before answering

Some questions deserve a quick answer. Others deserve a colleague who goes away and works it through properly. Claude can be both, and knowing when to switch is a genuine professional skill. In this lecture you will learn what Claude's thinking actually is, when it is worth the wait, how to frame a problem so the thinking is useful, and how to challenge the result.

### Why thinking changes answers

Why does thinking change the answer? Because a fast response is a bit like a quick opinion. It pattern matches to something familiar. That is perfect for most everyday tasks. But hard problems have constraints that interact. A budget cap here, a staffing limit there, a contract term that rules out an option. Those need working through, step by step. Thinking is Claude taking the long route on purpose, checking options against constraints before it commits. You trade some time for an answer that has actually been tested against your situation.

### What thinking is

Recent Claude models can reason before they respond. In the apps you switch on extended thinking. For developers, it is called adaptive thinking, where the model decides how much to think, steered by an effort level from low up to max. The newest models think by default and use effort to control depth. You will not see the raw internal reasoning in full. Depending on the product, you may see a summary or progress notes. That summary is handy for spotting wrong assumptions, but it is not a complete audit trail.

### Worth it vs not

When does it pay off? Multi constraint planning like budgets and schedules. Maths, pricing and forecasting logic. Comparing options against weighted criteria. Debugging a formula chain. Multi document analysis full of contradictions. Long agentic tasks. When is it wasted? Caption rewrites, short translations, formatting and simple lookups. The trade off is always the same, better answers on hard problems in exchange for time and usage.

### Thinking amplifies the problem statement

Here is the crucial point. Thinking amplifies the quality of your problem statement. Give it the goal and the decision. Give it all the data, not a summary. Give it the constraints, such as a minimum cash buffer, no hires before March, and fixed price contracts. Give it success criteria, maximise margin while keeping the risk of losing top clients low. And ask for an output shape, a comparison table, a recommendation, a numbered list of assumptions and a sensitivity check.

### Tell it what, not how

You do not need to tell current models how to think step by step. In fact, over prescriptive instructions can make results worse. Instead, tell Claude what a good answer must contain. The model is better at choosing its own reasoning path than you are at dictating it, but only you know what the decision requires.

### Challenge the result

When the answer arrives, read the assumptions list first, because wrong assumptions produce precise looking wrong answers. Check two or three key figures yourself, or ask Claude to recompute them with code. If something is wrong, correct it and rerun. Assumption four is wrong, we work twenty two days a month, not twenty. And ask for the counter case. Argue for the option you did not choose. It is the fastest way to stress test a recommendation.

### Simple example

A simple example. Ask Claude to schedule five client meetings next week, with these rules. No meetings on Friday afternoon, the Riyadh client only on Sunday to Thursday, at least one hour between meetings, and no more than two per day. With thinking off, you may get a neat looking schedule with a hidden clash. With thinking on, Claude checks each rule and produces a schedule that satisfies all of them, and tells you which rule was hardest to meet. That is the kind of task where thinking earns its time.

### Worked example: capacity planning in Lahore

A video editor in Lahore with clients in the UAE and the UK is offered a retainer of four edits a week. She switches thinking on and gives full data, current commitments, average edit hours, revision rounds and public holidays. Claude builds a capacity table and flags that one December week is infeasible because of UK revision rounds. The assumptions list reveals it treated a Pakistani public holiday as a working day. She corrects it, reruns, and negotiates a later start date with confidence. She saved the prompt as a capacity planning template. The next time a client offered a retainer, she updated her commitments and holiday calendar, reran it with thinking on, and read the assumptions first. It took ten minutes, and she replied to the client the same day with a start date she knew she could keep.

### Pitfalls

Three pitfalls. Using maximum thinking for everything and running out of usage by lunchtime. Expecting thinking to make up for missing data, because it cannot invent your costs. And treating a reasoning summary as proof. It shows direction, not verification. Your success signal is simple. The assumptions list regularly catches something you would have missed.

### Try this now

Try this now. Pick a real decision you face this month, a pricing change, a hiring plan, a budget split. Write the problem with the goal, full data, constraints and success criteria, and ask for a comparison table, a recommendation and a numbered list of assumptions. Run it once with thinking off and once with thinking on, in fresh chats. Compare the two. Which answer would you actually act on? Then pick the weakest assumption in the thinking version, correct it, and rerun. Notice how the recommendation changes. That is the habit that turns AI analysis into decisions you can stand behind.

### Watch me do it, part 1

Let me run the agency pricing decision. I switch extended thinking on and paste the template. Goal, recommend one of three pricing models for our six person content agency for the first quarter. Data, I paste our costs, the client list with monthly fees, hours per client and churn history. Constraints, a minimum monthly cash buffer, no new hires before March, two clients on fixed price contracts until June. Criteria, maximise gross margin, keep the risk of losing our top three clients low, simple to explain. Output, a comparison table, a recommendation, every assumption numbered, and what would change the recommendation. The answer arrives after a short pause, with a table and eleven numbered assumptions.

### Watch me do it, part 2

I read the assumptions before the recommendation. Assumption four says twenty working days a month. We work twenty two. Assumption seven assumes two revision rounds per deliverable, but three of our contracts allow unlimited revisions. I correct both and ask Claude to rerun. The recommendation changes from model B to model C, because unlimited revisions make the hourly model riskier. Finally, I ask, argue for the option you did not choose. It makes a fair case for B if we cap revisions in new contracts. Now I have a recommendation, the reasoning, and a contract change to discuss with my partner.

### Recap and next step

Recap. Thinking is for hard, multi step problems. Feed it complete data, constraints and success criteria, tell it what a good answer contains, and challenge the assumptions. Your next step is in the lesson. Take one real decision this month, run the template with thinking off and then on, and compare the assumptions and the answer you would actually act on.

## Key takeaways

- Extended (adaptive) thinking trades time and usage for better answers on complex, multi-step problems; effort controls depth.
- Use it for planning, maths, trade-offs, multi-document analysis, debugging and long tasks; skip it for simple rewrites.
- Give goal, full data, constraints, success criteria and output shape; say what a good answer contains, not how to think.
- Read the assumptions first, recheck key numbers and treat reasoning summaries as direction, not proof.

## Try it

Take one decision you face this month. Use the thinking prompt template with deeper thinking on and off, and compare the two recommendations and their assumptions.

- [Previous: Agentic tasks: multi-step work, Claude in Chrome, Excel, PowerPoint and Slack](https://optimizeall.com/learn/mastering-claude/agentic-tasks-chrome-and-office)
- [Next: Writing and editing workflows that sound like you](https://optimizeall.com/learn/mastering-claude/writing-and-editing-workflows)
- [All lessons of Mastering Claude (Anthropic)](https://optimizeall.com/learn/mastering-claude)