Skip to main content
Extended Thinking on the /anthropic/v1/messages endpoint allows Claude models to deeply reason through complex tasks before producing answers. You control reasoning token expenditure via the thinking parameter object.

The thinking Parameter Rules

  • max_tokens Must Exceed Budget: max_tokens must always be strictly greater than budget_tokens to ensure sufficient room for text completion.
  • Minimum Budget Limit: The minimum value for budget_tokens is 1024 tokens.
  • Response Block Structure: Responses emit an ordered block list where type: "thinking" precedes type: "text".

Next Steps