Table of Contents
What this lesson is about
Extended thinking is a mode in which Claude reasons through a problem step by step before giving you its answer — like watching a careful thinker work through something on paper rather than responding off the top of their head. This lesson explains when that deliberate process produces meaningfully better results, when it is wasteful, and how to control it so you get the quality you need at a cost that makes sense.
Core concept: the thoughtful pause before the answer
Imagine you ask a colleague a difficult question — say, whether your business should expand into a new market or consolidate what you already have. Two kinds of colleague exist. The first answers immediately, confidently, without visible hesitation. Sometimes that is exactly right. But for a genuinely complex question, you probably want the second kind: the person who says “give me a moment,” thinks it through out loud — “well, on one hand… but then again… the risk here is… what if we…” — and arrives at an answer that has been tested against its own objections before it reaches you.
Extended thinking makes Claude the second colleague. Instead of producing an instant response, Claude works through the problem internally, step by step, before composing its answer. You can read that working — the thinking trace — and see exactly how Claude reasoned its way to the conclusion. Nothing is hidden. The answer that emerges has been tested, weighed, and stress-checked before it arrives.
For simple questions, this is unnecessary. For genuinely difficult ones, it is often the difference between a good answer and the right answer.
How extended thinking differs from a normal response
In a standard Claude response, the process is effectively invisible. You send a message; Claude sends back an answer. Fast, direct, efficient — and perfectly appropriate for the vast majority of tasks.
Extended thinking adds an explicit reasoning phase before the final answer. During this phase, Claude:
- Identifies what is actually being asked and what makes it difficult
- Considers multiple approaches or perspectives
- Works through the implications of each
- Tests its emerging conclusions against potential objections
- Decides which answer is best supported by the reasoning
Only then does it write the response you receive.
The result is an answer with more depth, more nuance, and fewer overlooked angles. The trade-off is time and cost — extended thinking uses significantly more tokens (the unit used to measure text processed, roughly three-quarters of a word) than a standard response.
When to use extended thinking
Extended thinking earns its cost when the question itself has genuine complexity — when there is no single obvious right answer, when multiple considerations need to be weighed against each other, or when the consequences of getting it wrong are meaningful.
Strong candidates for extended thinking
- Strategic decisions with significant trade-offs — expanding vs consolidating, pricing changes, supplier choices
- Ambiguous problems where the right approach is not obvious from the surface
- Multi-step reasoning where each step depends on the previous one being correct
- Situations where you need Claude to surface risks or objections you have not thought of
- Analysis of conflicting information or data that seems to point in different directions
- Drafting complex documents where structure and argument need to be thought through before writing begins
When NOT to use extended thinking
| Task | Use thinking? | Why |
|---|---|---|
| Reformatting a table | No | There is nothing to reason about — just do it |
| Fixing a spelling error | No | Single-answer task — thinking wastes tokens |
| Writing a short bio | No | Straightforward execution task |
| Translating a paragraph | No | No ambiguity to resolve |
| Choosing between two pricing strategies | Yes | Multiple trade-offs requiring careful weighing |
| Diagnosing why a campaign is underperforming | Yes | Ambiguous problem with several possible causes |
| Writing a risk assessment | Yes | Requires identifying non-obvious risks |
| Drafting a policy document | Yes | Structure and argument need deliberate thought |
| Answering “what day is it?” | No | Factual lookup — no reasoning required |
| Deciding whether to enter a new market | Yes | Complex, high-stakes, multi-variable decision |
The test is simple: would a thoughtful human expert pause and reason carefully before answering this, or would they respond immediately? If they would pause, extended thinking is likely worth it. If they would not, it is not.
How to enable extended thinking
You can enable extended thinking in a Claude Code session by including a specific instruction in your prompt, or by using the thinking flag when available in your setup.
The clearest way to trigger it in a conversation is to ask explicitly:
Please use extended thinking for this. I want you to reason through the
problem carefully before giving me your answer — take the time you need.
When working via the API (the programming interface that developers use to connect Claude to their own tools and systems), extended thinking is enabled by adding a thinking parameter to the request. You do not need to know how to do this yourself — if you are using Claude Code directly, the explicit instruction in your prompt is sufficient.
Setting a thinking token budget
A thinking budget is the maximum number of tokens you allow Claude to use during its reasoning phase. This is how you control cost. A larger budget allows deeper, more thorough reasoning. A smaller budget limits the reasoning phase — useful when you want some deliberation without full depth.
Think of it like this: a budget of 1 000 thinking tokens is a five-minute think. A budget of 10 000 is an hour in the library. The right budget depends on the complexity of your question and how much you are willing to spend on the reasoning phase alone.
In practice, you can guide the budget in plain English:
Please think through this carefully, but keep the reasoning reasonably concise.
I do not need an exhaustive analysis — just thorough enough to catch the main risks.
Or, for a more detailed exploration:
Take as much thinking space as you need on this — I want a thorough analysis
that covers all the main angles before you give me your recommendation.
How to read the thinking trace
The thinking trace is the visible record of Claude’s reasoning — everything it considered before composing the final answer. It appears before the response itself, usually in a distinct section labelled something like “Thinking” or presented in a collapsible block.
What to look for in a thinking trace
When you read a thinking trace, you are looking for:
The problem restatement — Did Claude understand what you were actually asking? If the trace shows Claude framing a different question than the one you intended, that is useful information before you read the answer.
The options considered — What alternatives did Claude weigh? Seeing what was not chosen, and why, is often as valuable as the recommendation itself.
The objections raised and resolved — Did Claude identify weaknesses in its own emerging conclusion and test them? A good thinking trace will argue against itself before settling on an answer.
Assumptions made explicit — Extended thinking tends to surface assumptions that a quick answer would leave buried. If Claude assumed something important about your situation, it will often name it in the trace — which gives you a chance to correct it.
The point of decision — Where in the trace did Claude settle on its conclusion? What was the deciding factor? This tells you how confident the answer is and what would change it.
You do not need to read every thinking trace in full. For straightforward questions, skim for the conclusion and the main reason. For high-stakes decisions, read carefully — the trace is where the real analytical value lives.
Real example 1: a strategic business decision
Here is a situation where extended thinking is clearly appropriate — choosing between two meaningfully different directions for a business.
I want you to use extended thinking for this question.
My handmade ceramics business (Thandi's Ceramics) has grown steadily for
two years. I now have two realistic options for the next phase:
Option A: Open a small physical studio and retail space in Cape Town's
CBD — higher overheads, direct customer contact, brand visibility,
but significant fixed monthly costs.
Option B: Expand the online store aggressively — invest in photography,
paid social media advertising, and a PR push targeting lifestyle
publications — lower fixed costs, higher variable spend, no physical presence.
I currently turn over R480 000 per year, my margins are around 38%, and
I have R120 000 available to invest. I work alone with one part-time assistant.
Think through both options carefully and give me your honest recommendation
with your reasoning.
When extended thinking is enabled for a prompt like this, the thinking trace will typically:
- Restate the decision to confirm understanding
- Work through the financial implications of each option (fixed vs variable costs, break-even calculations, cash reserve implications)
- Consider non-financial factors (personal capacity, risk tolerance, brand positioning)
- Identify what would need to be true for each option to succeed
- Raise risks that were not mentioned in the prompt (lease lock-in periods, seasonality of foot traffic, algorithm changes affecting social reach)
- Weigh the options against each other given the specific constraints provided
- Arrive at a reasoned recommendation with clear conditions attached
The final answer you receive is the distilled conclusion of all of that reasoning — more reliable, more nuanced, and more aware of its own limitations than an instant response would be.
Real example 2: a simple task where extended thinking wastes money
Here is the same principle applied the wrong way — enabling extended thinking for a task that has no reasoning requirement.
With extended thinking enabled (wasteful):
Please use extended thinking to reformat the following list into a table
with two columns — item name on the left, price on the right.
Speckled mug — R285
Sage green bowl — R520
Off-white side plates (set of 4) — R680
Extended thinking would reason through something like: “The user wants a table. The data has two fields per row. I should use markdown table syntax. The headers should probably be ‘Item’ and ‘Price’…” — none of which requires deliberation. The final table would be identical to the one produced without thinking. The only difference is the token cost.
Without extended thinking (correct approach):
Please reformat the following list into a table with two columns —
item name on the left, price on the right.
Speckled mug — R285
Sage green bowl — R520
Off-white side plates (set of 4) — R680
Same result. A fraction of the cost. No thinking required.
Cost comparison: thinking on vs thinking off
The numbers below are illustrative — actual costs vary depending on the model, your usage plan, and the length of the specific prompt and response. The ratios, however, are realistic.
| Task type | Thinking off | Thinking on | Cost increase | Worth it? |
|---|---|---|---|---|
| Reformat a list into a table | ~300 tokens | ~2 500 tokens | ~8× | No |
| Write a short product bio | ~400 tokens | ~3 000 tokens | ~7.5× | No |
| Choose between two pricing strategies | ~800 tokens | ~6 000 tokens | ~7.5× | Yes |
| Diagnose why a campaign underperformed | ~1 000 tokens | ~8 000 tokens | ~8× | Yes |
| Draft a supplier contract clause | ~900 tokens | ~7 000 tokens | ~7.7× | Depends on stakes |
| Answer a factual question | ~200 tokens | ~2 000 tokens | ~10× | No |
The cost multiplier is fairly consistent across task types — extended thinking reliably costs significantly more. What changes is the value delivered. For tasks where deeper reasoning produces a meaningfully better answer, that cost is an investment. For tasks where there is nothing to reason about, it is simply waste.
Practical Exercise
In this exercise you will run the same question twice — once with extended thinking and once without — and compare the quality and cost of the two responses.
a. Choose a genuine business decision or strategic question you are currently facing — something with at least two plausible options and real trade-offs. Write a clear, detailed prompt that describes your situation, the options, and any relevant constraints (budget, time, capacity). Then send it to Claude with extended thinking enabled:
Please use extended thinking for this question. I want to see your
reasoning before you give me your recommendation.
[Your question here — describe the situation, the options,
and any relevant numbers or constraints]
When you receive the response, read the thinking trace first. Note: How did Claude frame the problem? What risks or considerations did it raise that you had not thought of? Did it make any assumptions you would want to correct?
b. Now send the same question again in a new session, without asking for extended thinking:
[The same question, without any instruction about thinking]
Compare the two responses. Is the recommendation the same? Is the reasoning in the standard response as thorough as the thinking trace? Did the extended thinking version surface anything genuinely useful that the standard version missed?
c. Now choose a simple formatting or editing task — something with a single obvious correct output. Send it with extended thinking enabled and note the response, then send it again without:
Please use extended thinking to correct the grammar in the following
paragraph and fix any punctuation errors.
[A short paragraph with a few deliberate errors]
Compare the outputs. Are they different in any meaningful way? What does this tell you about when extended thinking earns its cost?
Common problems and how to fix them
The thinking trace is very long and hard to parse
A long trace is not a problem in itself — it means Claude reasoned thoroughly. But you do not always need to read every line. Skim for the section where Claude settles on its conclusion (often signalled by phrases like “on balance,” “given these factors,” or “the stronger case is”) and read backward from there to understand the deciding factors. For high-stakes decisions, reading in full is worthwhile — treat it the way you would read a consultant’s working notes.
Extended thinking does not seem to produce better answers than normal mode
This usually means extended thinking was applied to a task that did not require it — a formatting job, a factual question, a single-answer task. There is no amount of reasoning that improves a correctly formatted table. Reserve extended thinking for genuinely ambiguous or multi-variable problems, and you will see a clear quality difference.
The cost feels too high for the complexity of the answer
If extended thinking consistently costs more than it seems to deliver, the most likely cause is that the thinking budget is too large for the questions you are asking. Try explicitly asking Claude to keep its reasoning concise: “Please think through this carefully but keep the reasoning brief — I do not need exhaustive analysis.” This encourages efficient reasoning rather than extensive exploration for its own sake.
The thinking trace seems to reason its way to a wrong conclusion
Read the trace to find where the reasoning went astray — it is usually an incorrect assumption Claude made about your situation. Correct the assumption explicitly in a follow-up: “In your reasoning you assumed [X], but actually [Y] — can you reconsider your recommendation with that in mind?” Because the trace makes the reasoning visible, you can pinpoint the error precisely rather than just disagreeing with the conclusion.
Extended thinking takes much longer to respond
This is expected. Deliberate reasoning takes time — that is the point. For time-sensitive situations where you need a quick answer, do not use extended thinking. For decisions where quality matters more than speed, the wait is the appropriate trade-off. If the delay feels excessive for the complexity of your question, try asking for more concise reasoning, as above.
What you have learned in this lesson
- Extended thinking is a mode in which Claude reasons through a problem step by step — working out loud before composing its answer — rather than responding instantly
- The thinking trace is the visible record of that reasoning, which you can read to see how Claude framed the problem, what options it considered, what risks it identified, and how it reached its conclusion
- Extended thinking is worth using for genuinely complex decisions, ambiguous problems, multi-step reasoning, and high-stakes situations where a more considered answer has real value
- It is wasteful — and produces no better output — for simple tasks, formatting work, factual questions, and any task with a single obvious correct answer
- A thinking budget controls the depth of reasoning and the associated cost; you can guide it in plain English by asking for concise or thorough reasoning
- Extended thinking reliably costs significantly more tokens than a standard response — roughly seven to ten times more for similar prompts — so the value of the question must justify the cost
- Reading the thinking trace critically — checking assumptions, noting what was considered and rejected, identifying the deciding factor — is where much of the analytical value of extended thinking is found