APIIntermediateClaude API
Turn on reasoning only where it changes the answer
Thinking tokens are billed as output: on classification or extraction you pay extra without gaining accuracy.
Choosing the model and enabling reasoning are separate decisions, and the second is made per request. On current models reasoning is adaptive and is tuned with an effort setting; budget_tokens is deprecated and returns a 400 error on the newest models. Check the documentation for your model's defaults and parameter names before implementing.
Example
Task → decision: - Classifying 50,000 tickets: reasoning off. - Planning a refactor with dependent steps: on, with matching effort. - Lookup or format conversion: off.