The meter used to be read before inference and incremented after it. Two batches arriving together at the edge of an allowance both read a number under the line, both ran, and both counted. On a hard ceiling, that meant the ceiling did not hold.
A call now reserves its record count in a single conditional update on the meter row before any inference runs. The reservation is refused if it would cross the ceiling, and it is released if the call fails afterwards, so a failed request stays free. The grace band on paid plans works as before: the reservation is refused only at the backstop. The Free plan’s agent loop runs on the same counter.
Two bounds on the assistant land with this. A message is capped at 8,000 characters and the page context it carries at 16 KB; over either, the request is refused before anything is stored. And the assistant’s per-user budget is keyed on the signed-in user rather than the caller’s network address.