Claude Extended Thinking: When and How to Use It

Claude extended thinking is a mode where the model works through a reasoning process before producing its answer. That reasoning appears as a separate, collapsible block — you can read the steps the model took, not just the conclusion. It is most useful for tasks that require several logical moves: complex analysis, multi-step calculations, or problems where the order of reasoning matters.

Enabling it does not guarantee a better answer on every task. Short factual questions, creative writing and simple summaries rarely benefit from the extra reasoning pass. The model uses more tokens and takes longer, so use it when the problem genuinely warrants it. Keeping the problem description and the accepted conclusions in a tool like fenbs means each session starts from the agreed position rather than repeating the reasoning.

What extended thinking actually does

Standard Claude responses are generated in a single pass. With extended thinking enabled, the model first produces a chain of reasoning — exploring the problem, checking assumptions, considering alternatives — and then writes the visible answer. The reasoning block is shown to you but is separate from the final response.

This matters because it changes how errors surface. In a standard response, a wrong assumption can carry quietly through to the conclusion. In extended thinking, you can often see where the model went wrong and correct the reasoning before relying on the output.

Which tasks benefit most

• Multi-step reasoning problems: logic puzzles, mathematical proofs, legal argument chains, financial modeling.
• Decision analysis where the model should weigh several competing factors before recommending.
• Debugging complex code where the cause is not obvious and the model needs to trace through several possibilities.
• Research synthesis where contradictory sources need to be reconciled before drawing a conclusion.
• Tasks where you want to audit the reasoning, not just accept the answer.

If you are preserving the problem brief and the decision rationale for later use, tools like fenbs let you record what the model concluded and why, so the next session (or a different AI app) has the agreed answer and the reasoning behind it.

How to enable extended thinking in Claude

Access depends on which Claude surface you use. The feature is not available on every plan or in every interface.

In claude.ai

In the web app, extended thinking is available on plans that include the relevant Claude models. Look for a thinking control in the message input area. Anthropic's help center is the right place to check current plan availability, because access can change as models are updated.

In Claude Code

Claude Code uses extended thinking automatically on harder problems when the model determines it is warranted. You can also explicitly request it. Check the Claude Code documentation for current options. Claude Code Projects added shared memory in September 2026, which means the project context is available to every session, reducing the need to re-explain the problem in the thinking pass.

Via the Anthropic API

The Anthropic API exposes extended thinking through a thinking parameter in the request. You set a budget_tokens value that limits how many tokens the model can spend on reasoning before writing the response. Anthropic's API documentation describes the current parameters; the interface has changed between API versions, so read the version you are using.

Reading and verifying the reasoning block

The extended thinking block shows the model's working, but it is not a verified proof. The model can make mistakes in the reasoning chain and still arrive at a plausible-sounding conclusion. Read the reasoning with the same skepticism you would apply to the answer itself.

Useful things to check: did the model identify the right question, does it use the facts you gave it or does it introduce assumptions, and does its conclusion follow from its reasoning? If you spot a wrong turn, you can correct it in the next message rather than accepting the whole answer.

Extended thinking and context limits

The reasoning tokens count toward the session context. A very long thinking chain can consume a significant portion of the available context window, leaving less room for the conversation and the source material. If you are working with long documents, monitor the context usage and break the problem into smaller pieces if needed.

For problems related to extended thinking not appearing or stopping unexpectedly, see the extended thinking not working guide.

How to do this with fenbs

1. Create a free account and add the project you are working on.
2. Write the problem brief as a task or a context note. Be specific: the model's reasoning quality depends on the quality of the description it receives.
3. Connect Claude to the project from /integrations. Ask it to read the context before starting the extended thinking session.
4. After the session, record what was decided in Decisions and rules. Include the key reasoning step that led to the conclusion.
5. Before you stop, ask Claude to save progress. The context and the decision are then available to the next session without restatement.