Max output tokens

Each model has a hard ceiling on response length, separate from and smaller than its context window, and API callers can set a lower per-request limit. When generation hits the limit, the response is cut off and the API flags the truncation via a finish or stop reason.

The parameter doubles as a cost control: it caps worst-case spend on the most expensive token type. Setting it slightly above the longest response you actually need protects against runaway generation without risking truncated answers.

Reasoning models complicate the calculation because hidden thinking tokens draw from the same output budget. A limit that comfortably fits the final answer can still truncate a response if the model spends most of the budget thinking, so reasoning modes generally need much larger output allowances.