Source: Claude API Docs — Thinking

Anthropic's reference for Claude thinking: modes per model, display, signatures, tool use, interleaving, caching, context-window rules.

Contents
  • What: official API documentation page
  • Publisher: Anthropic
  • Date: accessed 2026-11-14 (living document; API version header 2026-10-21 in examples)
  • Original: Thinking

Summary

The mode-neutral reference for how Claude’s thinking works on the Messages API. It gives a per-model table of what each thinking value does (adaptive, enabled + budget_tokens, between_tools, disabled), explains thinking blocks and their encrypted signature, the display option, the rules for passing thinking blocks back during tool use, interleaved thinking, which models keep prior-turn thinking in context, prompt-caching interactions, and feature incompatibilities.

Key claims

  • Thinking arrives as thinking content blocks before text blocks. Each carries a signature: an encrypted copy of the full reasoning that must be passed back unchanged.
  • Visible thinking text is never the raw chain of thought. It is a summary produced by a different model. display: "summarized" returns it; display: "omitted" (default on Claude 5-generation models, Opus 4.7/4.8) returns an empty field. Billing is for the full thinking either way; omitting only cuts time-to-first-text-token.
  • On Claude 5-generation models and Claude Mythos Preview thinking is on by default as adaptive thinking; Opus 4.7/4.8 (and 4.6 and earlier) default to thinking off. enabled with budget_tokens returns 400 on Claude 4.7 and later (Mythos Preview still accepts it). Several models reject disabled (Opus 5.5, all Fable/Mythos); Opus 5 and Haiku 5.5 accept it only at effort high or below. Sonnet 5.5’s lowest setting is between_tools (up-front thinking off), also only at high or below.
  • effort (in output_config) controls how much work the whole response gets; max_tokens is the only hard ceiling on spend.
  • A tool-use loop is one assistant turn. Thinking blocks within that turn must be returned complete and unmodified (including redacted_thinking blocks); modified blocks give a 400. Toggling thinking mid-turn silently disables it for that request.
  • Manual extended thinking only allows tool_choice auto or none; adaptive thinking supports forced tool use, except on Opus 5.5, Sonnet 5.5 and Fable/Mythos 5.1 and 5.2, which reject forced tool use on every request (use auto with strict tools or structured outputs).
  • Interleaved thinking (reasoning between tool calls) is automatic with adaptive thinking; the thinking allocation can span the whole turn.
  • Keep-all models (Opus 4.5+, Sonnet 4.6+, Haiku 5.5, Fable/Mythos) keep prior turns’ thinking in context, billed as input; older models strip it once a non-tool-result user message arrives.
  • Thinking configuration and resolved effort are rendered into the prompt, so changing them invalidates cache breakpoints.
  • Opus 4.7+ and Claude 5-generation models reject non-default temperature/top_p/top_k (their top_p default is 0.99); from API version 2026-10-21 any sampling parameter returns a 400 on every model. Claude 4.6+ rejects assistant prefill. Max output is 128K on current models (300K via a batches beta on some). SDKs require streaming above 21,333 max_tokens.
  • A request that tries to extract internal reasoning into the response text may be refused on recent models (reasoning_extraction).

Feeds

Extended thinking · Reasoning models · Test-time compute

Suggest changes →