Study · Claude Course

Extended Thinking

Claude reasons before answering — higher accuracy on complex tasks, at the cost of thinking tokens + latency. Enable only **after** prompt optimization fails; use [[Prompt-Evaluation]] to decide.

type conceptstatus activeapi · reasoning · advanced

Key points

  • thinking_budget ≥ 1024; max_tokens must exceed it.
  • Response = a thinking block (reasoning + signature) + a text block. "Redacted" blocks are encrypted thinking flagged by safety.
Anthropic-API-Basics · Prompt-Evaluation

Sources

  • 2026-06-28-claude-course
Compiled from wiki/study/claude-course/Extended-Thinking.md · git is the source of truth