Your limit did not reset at midnight.
Consumer AI caps run on rolling windows, which is why they feel unpredictable.
A rolling window means your capacity returns gradually, message by message
The intuition people bring is a calendar quota: some allowance that refills at a fixed hour. Consumer AI products almost universally use a rolling window instead — capacity is measured over the last N hours from *now*, continuously. The practical difference is large. Under a calendar quota, using everything at 9am means waiting until midnight. Under a rolling window, the messages you sent at 9am stop counting against you at 9am plus N, so capacity trickles back in the same pattern you spent it. This is why the limit feels like it moves, why waiting "a bit" sometimes works and sometimes does not, and why heavy use in a burst is punished more than the same volume spread out.
- Rolling window
- Usage measured over the last N hours from now, recalculated continuously.
- Why it feels random
- Capacity returns in the pattern you spent it, not all at once at a fixed time.
- What counts
- Usually tokens, not messages — so long conversations consume far faster.
- Context re-billing
- Each turn re-sends the whole thread, so late messages cost more than early ones.
- Model tier matters
- Frontier models draw on a smaller allowance than fast ones on the same plan.
- The reliable fix
- Start a new conversation. It resets the context you are paying for each turn.
FAQ
Because most caps count tokens rather than messages, and every turn re-sends the entire history. Your fiftieth message in a thread might carry twenty times the tokens of your first, so it consumes twenty times the allowance. Starting a fresh conversation for a new topic is the single most effective habit for staying under a cap.
Under a rolling window there is no single moment — capacity returns progressively as your oldest usage ages out of the window. Waiting the full window length from your heaviest burst is the reliable answer; waiting a few minutes sometimes works if you were only just over.
Often, because tiers usually draw on separate allowances. A faster or smaller model on the same plan frequently still has capacity when the frontier model does not, and for a lot of tasks the difference in output is smaller than people expect.
Related but not the same. API rate limits are typically expressed as requests and tokens per minute and are about protecting infrastructure; consumer usage caps are about plan economics and run over hours. API limits also return a specific retry-after header, which consumer products generally do not surface.