Extended thinking
Extended thinking is now a legacy feature, so learn how adaptive thinking and effort work on hard problems and what the Thinking summary shows.
-
How Do You Get Claude to Work Through Math and Logic? Give Claude the whole problem, raise effort, and ask it to check its result against the numbers you gave. Level 3: WorkflowsDo
- State the full problem
- Raise effort
- Ask for a check against your numbers
Don't- Ask for a one-line answer
- Tell it to answer immediately
- Size max_tokens for the answer alone
-
How Do You See Claude's Thinking, and Can You Trust It? In the Claude apps, expand the Thinking section above a response. Level 3: Workflows
- 1Full thinkingbilled, never shown
- 2Summary you can expandcheck assumptions here
- 3The answerstill needs its own check
-
How Do You Tell Claude What a Right Answer Looks Like? List the conditions a correct answer must meet, in terms that can pass or fail, and ask Claude to verify against them before it finishes. Level 3: WorkflowsNothing to testRuns and reports
- "Make it good"
- Pass or fail conditions
- A test that runs
-
When Should You Turn Up Claude's Thinking? On current Claude models, you don't decide whether Claude thinks. Level 3: Workflows
- lowquick work you can check
- mediumdefault on Opus 5.5, Haiku 5.5
- highdefault on most other models
- xhighlong agentic and coding work
- maxhighest capability
This feature was renamed and changed
This lesson keeps its old address, but the feature has moved on. Extended thinking meant turning on a mode and setting a token budget. Anthropic now calls that a legacy feature, and Claude 4.7 and later models reject it. On current models, Claude decides for itself on each request whether to think, and you steer it with effort. In the apps, click the model name next to the send button, then Effort.
What you can actually control
You can't guarantee Claude will think on a given request, and no effort level promises it. You can raise effort, describe what a right answer must satisfy, and read the summary to catch a wrong turn. What you see is a summary, never the raw reasoning, and asking for the reasoning in the answer can be refused on current models. Higher effort takes longer and uses your plan's limits faster, so save it for work that repays it.
Is extended thinking still available in Claude?
As a manual setting with a token budget, only on older models. Anthropic calls it a legacy feature, and Claude 4.7 and later reject it. Current models use adaptive thinking, and some apps still label the toggle Extended on models where it can be switched.
What is adaptive thinking in Claude?
Adaptive thinking means Claude judges each request and decides whether to think and for how long. A simple question may get no thinking, while a tricky bug may get a lot. The effort setting shifts how readily and deeply it thinks.
What effort level should I use in Claude?
Start with the level marked Default in the menu. Anthropic's defaults are medium on Opus 5.5 and Haiku 5.5, and high on most other models. Raise it for hard problems where mistakes are costly, and lower it for quick, routine work.
Can I see Claude's thinking?
In the apps, expand the Thinking section above a response to see a summary of its approach. On the API, current models hide it by default until you set display to summarized. Either way it's a summary, not the raw reasoning.
Does thinking cost extra in Claude?
On the API, thinking tokens are billed as output tokens, and you pay for the full thinking rather than the summary you see. On a plan, Anthropic says higher effort reaches usage limits faster.