What Is Extended Thinking?

AI Summary

What it is: Extended thinking is a Claude feature that lets the model spend extra time reasoning through a problem before producing the final answer — trading speed for accuracy on hard tasks.

Who it’s for: Anyone using Claude for math, code, complex analysis, multi-step plans, or any task where the right answer matters more than the fastest answer.

Best if: You’re asking Claude to solve something where wrong answers are expensive — debugging tricky code, working through a financial model, planning a multi-step project.

Skip if: You’re doing a quick lookup or casual chat — extended thinking adds latency for marginal benefit on simple questions.

What is extended thinking?

Extended thinking (sometimes called “reasoning” or “think tokens”) is Anthropic’s name for letting Claude take more compute time per question before answering. Instead of producing an answer immediately, Claude generates an internal reasoning trace, evaluates it, and only then produces the final response. The reasoning trace is partially visible to the user in supported interfaces.

This is the same pattern OpenAI uses in its o-series models (o3 and o4-mini) and Google uses in Gemini’s “Deep Think” mode. The technical name across the industry is inference-time reasoning or test-time compute.

When does extended thinking help?

Three categories of task where it materially improves answer quality:

  • Math and quantitative reasoning. Multi-step calculations, financial modeling assumption checks, statistics.
  • Code that requires planning. Debugging a tricky bug, refactoring across multiple files, designing a system from requirements.
  • Complex multi-step plans. Project plans, decision trees, anything where the steps depend on each other.

It does not help much on quick lookups, casual chat, or tasks where the right answer is obvious from the prompt.

How do you turn extended thinking on?

In Claude.ai, extended thinking is available in Opus 4.7 and Sonnet 4.6. Toggle it on in the model picker before sending your message. In the Claude API, extended thinking is a request parameter on supported models; see Anthropic’s documentation.

Frequently asked questions

Does extended thinking cost more?
Yes. You pay for the additional “think tokens” Claude generates internally. On most paid plans this is included up to a quota; on the API it’s metered.

Can I see what Claude is thinking?
Partially. Claude.ai shows a summary of the reasoning trace in supported models. The full trace is not always exposed for safety and clarity reasons.

How does extended thinking compare to o3?
Same family of technique — both spend additional inference compute on internal reasoning. Different labs implement it differently. In practice, the model strength matters more than the label.

Get Smarter About AI Every Morning

Free daily newsletter — one story, one tool, one tip. Plain English, no jargon.

Free forever. Unsubscribe anytime.

Sources

You might also like

Two ways to go further

The AI Prompt Library

1,000+ ready-to-use prompts for Claude, ChatGPT, and Gemini. Stop staring at a blank box.

Get it for $39 →

2-Hour Live AI Crash Course

A private, beginner-friendly session across Claude, ChatGPT, Gemini, and the wider landscape.

Book for $125 →