Source: Claude API Docs — Steering Thinking

Anthropic's guide to adaptive thinking: effort levels, prompt steering, cache effects, cost control and billing of thinking tokens.

Contents
  • What: official API documentation page
  • Publisher: Anthropic
  • Date: accessed 2026-11-14 (living document)
  • Original: Steering thinking

Summary

Explains adaptive thinking: the model decides per request whether to think and how much. Effort is the main lever, with system-prompt and per-message wording as secondary levers. It also covers caching, cost bounds and what is billed.

Key claims

  • Effort levels: low, medium, high, xhigh, max. high is the default on most models; medium on Claude Opus 5.5 and Haiku 5.5. No level guarantees a thinking block on every request.
  • In tool loops, the first request after new user input usually carries most of the reasoning; follow-ups that only process tool results may skip thinking. Thinking per request tends to fall as conversations grow.
  • Prompt steering (“think carefully”, “respond directly”) works on some models but not reliably on others (Sonnet 5.5, Haiku 5.5, Fable 5.2); lowering effort is the calibrated lever.
  • Changing the top-level effort can invalidate the prompt cache; a beta per-message effort change preserves it.
  • Cost control: max_tokens is a hard per-request cap (not per turn in a tool loop); effort is soft. If you see stop_reason: "max_tokens", raise the cap or lower effort.
  • Billing: thinking tokens are output tokens; preserved prior-turn thinking is input; summary generation is free; billed output does not match visible tokens.
  • When thinking is active the API adds a specialized system prompt.

Feeds

Extended thinking · Test-time compute · Reasoning models

Suggest changes →