Extended thinking gives Claude a dedicated reasoning pass before each response. You trade some latency for significantly better answers on hard problems.

What it does

When extended thinking is on, Claude runs internal reasoning ahead of its visible response. You can toggle it mid-session with Alt+T. The reasoning itself isn't shown by default but can be surfaced for debugging. On complex architecture, subtle bugs, or multi-step plans, the quality jump is large.

When to use it

Design decisions with trade-offs to weigh.
Bugs that depend on subtle state.
Refactors spanning many files.
Code review where you want Claude to spot non-obvious issues.

Gotchas

Every turn gets slower with extended thinking on. Turn it off for rote edits.
Extended thinking consumes thinking tokens, which bill differently. Check your plan.
It's not a substitute for good prompts - garbage in still produces garbage out, just after longer reasoning.

Official docs: https://code.claude.com/docs/en/common-workflows.md#use-extended-thinking-thinking-mode