What is extended thinking?
Extended thinking is a Claude mode in which the model writes its reasoning in separate thinking blocks before the final answer, introduced by Anthropic with Claude 3.7 Sonnet in February 2025.
Anthropic launched extended thinking with Claude 3.7 Sonnet on 24 February 2025, describing a single model that could answer near-instantly or reason step by step first. API users set a thinking budget in tokens, and thinking tokens were billed as output.
The mechanism has since changed. Anthropic's docs now list manual extended thinking, with a fixed budget_tokens value, as a legacy feature: it is deprecated on Claude Opus 4.6 and Sonnet 4.6, and Claude 4.7 and later models reject it with an error. The replacement is adaptive thinking, where the model decides whether and how much to think and an effort setting controls depth. Teams still search for the old name, and many codebases still carry the old parameter.
How it works
Thinking blocks
The response contains thinking blocks, returned in summarised form, followed by the normal text answer.
Manual budgets (legacy)
budget_tokens had a minimum of 1,024 and had to be lower than max_tokens. Anthropic advised starting near the minimum for simple tasks and at 16,000 or more for complex ones.
Adaptive thinking (current)
Set thinking to adaptive and control depth through effort. At lower effort the model may skip thinking on easy inputs.
Billing
Thinking tokens are billed as output, count toward max_tokens and occupy the context window.
Extended thinking with tool use
When a tool result goes back to Claude, the thinking block that came with the tool call has to be returned with it, unmodified and with its signature, or the API returns an error. This is the detail most often broken by home-made agent loops that rebuild message history.
Interleaved thinking lets the model reason between tool calls inside one turn. It is automatic with adaptive thinking; on older models it needed a beta header. Changing the thinking budget or mode mid-conversation also invalidates prompt cache breakpoints, so the setting should stay stable within a cached session.
What it is often confused with
The think tool
Anthropic's think tool (March 2025) is a no-op tool the model can call mid-trajectory to stop and reason, aimed at policy-heavy tool chains. It predates interleaved thinking and solves a narrower problem.
Chain-of-thought prompting
Asking the model in the prompt to reason step by step. Anthropic's current guidance is to rely on built-in thinking on its newest models, and notes that a prompt asking for reasoning in thinking tags may be declined.
A reasoning model
The broader class. Extended thinking is Anthropic's implementation of it.
How Motiontic's engineers apply it
When moving a client's agent to a current Claude model, Motiontic's engineers remove fixed budgets, set effort per step from eval results, check that the loop returns thinking blocks intact, and compare cost and quality before and after on the regression suite.
Manual extended thinking compared with adaptive thinking
| Manual extended thinking | Adaptive thinking | |
|---|---|---|
| Status | Legacy; rejected by Claude 4.7 and later | Current default approach |
| Control | budget_tokens, minimum 1,024 | Effort level |
| Thinks on every request | Yes | Only when the model judges it useful |
| Between tool calls | Needed a beta header on older models | Automatic |
| Fits | Older models; strict latency or cost control | Current Claude models |
Common questions
What is extended thinking in Claude?
A mode in which Claude reasons in separate thinking blocks before giving its answer. It was introduced with Claude 3.7 Sonnet in February 2025 and has been succeeded on newer models by adaptive thinking.
Is extended thinking deprecated?
Manual extended thinking with budget_tokens is. Anthropic's docs list it as legacy, deprecated on the 4.6 models and rejected by Claude 4.7 and later, which use adaptive thinking with an effort setting.
Are thinking tokens billed?
Yes, as output tokens. They also count toward the maximum output for the request and toward the context window.
What is the difference between extended thinking and the think tool?
Extended thinking is reasoning the model does before or between its responses. The think tool is a tool the model calls to pause and reason mid-task. Interleaved adaptive thinking now covers most of what the think tool was used for.
Related
Reasoning model
A model that thinks before it answers, at a price in tokens and time.
Think tool
A tool that does nothing, so the model has somewhere to stop and think between actions.
Chain-of-thought prompting
Asking the model to show its working, now mostly built into the model itself.
Prompt caching
Paying less for the part of the prompt that does not change between calls.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.