AI Summary API: Answers Are No Longer Cut Short

Ask the AI Summary API a short question, such as a request for a small PHP function, and the answer used to stop a few paragraphs in, sometimes halfway through a line of code, with a full stop glued to the end so it looked finished. The Document Reader and AI Hints never showed it, because the pages they send are long. That limit is gone.

01 What changed

• The answer is sized by the answer. It used to be capped at a multiple of the length of the prompt, which gave a one-line instruction room for 256 tokens. The model now stops when it has finished, within the room left in its window, up to 8,192 tokens.
• A cut answer says it was cut. In the rare case an answer does reach the limit, it ends with [Answer truncated: output limit reached] instead of a full stop that was never written by the model.
• Nothing is rewritten. An answer that ends on a code block, a list or a table is delivered as the model wrote it, streamed or not.
• Cut answers are not cached. Answers are cached so a repeated question costs nothing; one that hit the limit is never stored, so asking again produces a fresh answer.

No change is needed on your side: the endpoint and its parameters stay the same.