When to Use Extended Thinking

Extended thinking is one of Claude's most powerful — and most misunderstood — capabilities. It's not a magic performance boost for every task. Used in the right situations, it produces dramatically better results. Used in the wrong ones, it just increases cost and latency with no benefit. This lesson explains what extended thinking actually is, where it helps, and how to think about the cost tradeoffs.

---

What Extended Thinking Actually Is

When you enable extended thinking, Claude gains access to a scratchpad before composing its response. During this phase, Claude works through the problem — exploring approaches, checking reasoning, reconsidering assumptions, and organizing its thoughts — before producing the final answer you receive.

These internal steps are called thinking tokens. They are billed as output tokens and count toward max_tokens, but the raw reasoning is not shown to the end user. What you get back is the finished response, informed by that internal reasoning process.

Think of it like the difference between asking someone a hard question and expecting an instant answer versus giving them five minutes to think on paper before they respond. The thinking process itself isn't the output — it's the mechanism that makes the output better.

---

How You Control Thinking: Adaptive Thinking and Effort

On the latest Claude models, thinking is adaptive. Claude decides on each request whether to think and how much. You do not set a token budget. You steer the depth with the effort setting.

{
  "model": "claude-opus-5-5",
  "max_tokens": 16000,
  "thinking": {
    "type": "adaptive"
  },
  "output_config": {
    "effort": "high"
  },
  "messages": [...]
}

How it works:

• effort goes inside output_config. The levels are low, medium, high, xhigh, and max
• Higher effort means deeper reasoning, more tokens, and a slower response
• At lower effort, Claude may skip thinking on easy requests
• The default level depends on the model, so set effort yourself when depth matters
• Thinking counts toward max_tokens. Leave room for both the reasoning and the answer

On the newest models, thinking is on by default. Some of them do not let you turn it off, and effort is the only control. Check the models overview for the model you use.

Older models used a fixed budget. Before adaptive thinking, you turned thinking on with "thinking": {"type": "enabled", "budget_tokens": 10000}. The budget had to be at least 1,024 tokens and less than max_tokens. That form is deprecated on the Claude 4.6 models. Claude 4.7 and later models reject it with a 400 error. You will still see it in older code and tutorials. When you move that code to a current model, remove budget_tokens, set the thinking type to adaptive, and pick an effort level.

In this course, "extended thinking" means this reasoning phase in general, whichever way you control it.

---

When Extended Thinking Helps

Extended thinking improves performance on tasks that require holding multiple considerations in mind simultaneously, checking work against constraints, or exploring a problem space before committing to an approach.

Complex mathematical and quantitative reasoning Multi-step calculations, proofs, or optimization problems where errors compound across steps. Claude can verify intermediate results before proceeding.

Multi-step logical problems Puzzles, deductive reasoning chains, or inference tasks where the answer depends on correctly following a sequence of logical steps.

Code architecture decisions Designing a system from scratch, choosing between architectural approaches, or diagnosing performance issues that require reasoning about the entire call chain.

Legal and financial analysis Applying multiple overlapping rules, regulations, or constraints to a specific situation — the kind of analysis where a wrong intermediate step invalidates the entire conclusion.

Debugging complex issues Tracing a bug through multiple layers of code or data pipelines where understanding root cause requires ruling out many candidate explanations.

Strategic planning under constraints Any task with multiple competing requirements that need to be weighed against each other before arriving at a recommendation.

---

When Extended Thinking Doesn't Help

For many tasks, extended thinking adds cost and latency with no quality improvement. The internal reasoning process doesn't change the output when the task doesn't require it.

Simple factual Q&A — "What is the capital of France?" Thinking won't change the answer.

Creative writing — Writing a poem or short story is not improved by internal deliberation. Claude's creative output is already a generative process.

Translation — Translating text between languages is a pattern-matching task that doesn't benefit from extended reasoning.

Summarization of short content — Summarizing a paragraph or two doesn't require pre-reasoning.

Straightforward instructions — "Format this list alphabetically" or "Make this email more formal" are direct edits that don't require extended thinking.

High-volume, low-complexity tasks — If you're running thousands of classification requests, the cost of thinking tokens adds up quickly for no measurable benefit.

A useful heuristic: if a task could be done well by a thoughtful person in under 30 seconds, extended thinking won't help. If the task would benefit from someone working through it carefully on a whiteboard, it probably will.

---

Cost Implications

Thinking tokens count toward output token usage and are billed accordingly. You pay for the full reasoning even when the API does not return the thinking text. A request at high effort can cost much more than the same request at low effort.

Practical guidelines:

• Don't run every request at high effort
• Identify the small subset of your use case that genuinely needs deep reasoning
• For interactive applications, set expectations about latency. Deeper thinking takes longer
• Test with the lowest effort level that produces acceptable quality before you raise it

---

Try It: A Problem Suited for Extended Thinking

The prompt below presents a constraint optimization problem — the kind of task that benefits most from extended thinking. Try it as written, then experiment with simplifying the problem to see when the extra reasoning starts to matter.

---

Key Takeaways

• Extended thinking gives Claude a reasoning scratchpad before producing its response — thinking tokens are not shown to users
• It helps most for tasks requiring multi-step reasoning, constraint satisfaction, complex debugging, and strategic analysis
• It adds no value for simple Q&A, creative tasks, translation, or straightforward instructions
• On current models, thinking is adaptive. Steer depth with the effort setting and raise it only when output quality is insufficient
• The budget_tokens form is for older models only. Claude 4.7 and later models reject it
• Thinking tokens are billed as output tokens — enable selectively for tasks that genuinely require it