The maximum number of tokens a model is permitted to generate in a single response, set by configuration rather than by the model running out of things to say.

Presented as a technical constant, when it is a dial the vendor sets, can lower at any time, and is under no obligation to publish or to announce when it moves.

• A Dial Described as a Constant: The limit is chosen, and the phrasing implies it was discovered.
• Silent Adjustment: Nothing notifies you if it is reduced, and the shorter answers arrive looking perfectly normal.
• House-Cleaning Reality Check: Like the same-priced chocolate bar getting quietly thinner each year, with the wrapper unchanged.

The Menu (2020s) where a precisely measured, minuscule portion arrives at an extreme fixed price, and any diner who remarks on the size is told the shortfall lies in their own understanding.

“Responses are bounded by the output token limit configured for this service tier.”

“Answers are capped at roughly two pages, and the cap can be lowered without anyone telling us.”