Extended thinking

Overview

• Extended thinking allows Claude to reason about queries before generating final responses
• Displayed as a separate thinking process that users can optionally view
• Enables tackling complex tasks with greater accuracy

Trade-offs

• Cost: You are charged for tokens generated during the thinking phase
• Latency: The thinking phase takes additional time to complete

When to Enable Extended Thinking

• Use prompt evaluations to determine if extended thinking is needed
• Enable it only after spending effort improving your prompt and accuracy is still insufficient
• Decision should be based on eval results, not used by default

Response Structure with Extended Thinking

• Normal responses contain: user message (text block) → assistant message (text block)
• With extended thinking enabled, responses contain: thinking block + text block
• Thinking block contains the text generated during Claude's reasoning process

Thinking Block Signature

• Each thinking block includes a cryptographic signature
• Signature ensures the thinking text has not been modified
• Claude relies heavily on thinking text during response generation
• Developers cannot modify thinking text to prevent unsafe steering

Redacted Thinking Blocks

• Sometimes Claude flags thinking text through internal safety systems
• Response contains a redacted thinking block with encrypted content instead
• Redacted content preserves context when sending the message back to Claude
• Useful for testing application handling of unexpected responses

Implementation Parameters

• thinking: Boolean parameter (default false) to enable extended thinking
• thinking_budget: Number of tokens allocated for thinking phase (minimum 1024 tokens)
• max_tokens: Must be greater than thinking_budget (e. g. , if budget is 1024, max_tokens must be at least 1025)
• Best practice: Set max_tokens significantly larger than thinking_budget to allow adequate response generation

Testing Redacted Thinking

• Use special test string: "anthropic magic string triggered redacted thinking" followed by special characters
• Sending this exact string guarantees a redacted thinking block response
• Useful for verifying application handles redacted blocks without crashing

Takeaways

• Extended thinking enables Claude to reason through complex queries before responding, with costs for thinking tokens and added latency
• Enable extended thinking only after improving prompts and eval results show accuracy is still insufficient—not by default
• Responses with extended thinking contain a thinking block (Claude's reasoning) plus a text block (final response), with thinking blocks cryptographically signed to prevent modification
• Set thinking_budget (minimum 1024 tokens) and ensure max_tokens is significantly larger to allow adequate response generation after the thinking phase
• Redacted thinking blocks may appear when internal safety systems flag content; they preserve context for continued conversation without exposing the flagged reasoning