PRODUCTION SCENARIO
A developer configures Claude Sonnet 4.5 for algorithmic verification. The Messages API request sets thinking to type enabled with budget_tokens 2048 and max_tokens to 2000. Every call returns HTTP 400 before any tokens are generated.
Why does the request fail validation?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: max_tokens must be strictly greater than budget_tokens on this model
Thinking tokens count toward max_tokens, so budget_tokens must be less than it or nothing is left for the answer; here 2048 exceeds 2000. 1,024 is the minimum budget, not a threshold needing a beta header, and the interleaved-thinking header only relaxes this rule inside one assistant turn. A non-default temperature is indeed rejected while thinking is on for this model, but the request never sets one, so that rule is not what fails here. Thinking is a top-level parameter; output_config carries effort on adaptive-thinking models.