Source: Claude API Docs — Extended Thinking (manual budget)

Anthropic's reference for the legacy budget_tokens thinking mode: budget rules, interleaving beta header, caching, migration to adaptive thinking.

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

Summary

Covers manual extended thinking, thinking: {type: "enabled", budget_tokens: N}, now marked legacy. It sets out the budget constraints, tuning advice, how interleaved thinking is switched on per model generation, the caching consequence of changing the budget, and how to migrate to adaptive thinking with effort.

Key claims

  • Deprecated: Claude 4.7 and later reject type: "enabled" with a 400. Still the only mode on Opus 4.5, Sonnet 4.5 and Haiku 4.5; deprecated on Opus/Sonnet 4.6.
  • budget_tokens minimum is 1,024 and must be below max_tokens (thinking counts toward it), except with interleaved thinking, where the budget spans every thinking block in the turn.
  • The budget is a target, not a strict cap; max_tokens is the hard ceiling.
  • Tuning: start near the minimum for simple tasks, 16K+ for complex ones; returns diminish. Above 32K, use batch processing to avoid timeouts.
  • usage.output_tokens_details.thinking_tokens reports billed reasoning tokens (final message_delta when streaming).
  • Interleaved thinking in manual mode needs the interleaved-thinking-2025-05-14 beta header on Claude 4/4.5 models; Opus 4.6 manual mode cannot interleave; Haiku 4.5 does not interleave.
  • Changing budget_tokens invalidates cache breakpoints because the value is rendered into the prompt.
  • Migration: drop budget_tokens, set type: "adaptive", control depth with output_config.effort. Behaviour changes: a fixed budget thinks every request; adaptive may skip thinking on easy inputs.

Feeds

Extended thinking · Reasoning models

Suggest changes →