The services that generate images limit how many requests they will take in a given minute. Reaching that limit is ordinary and temporary: the provider refuses the request, states how long to wait — usually a matter of seconds — and accepts the next one.
Until now the platform could read that refusal as something else entirely. The provider answers a short-term limit with the same sentence it uses when there is a problem with the account behind the service, and the platform acted on the sentence rather than on the waiting time offered alongside it. A generation that only needed a few seconds could therefore be marked as failed on its very first attempt, with no retry.
The platform now reads the waiting time the provider supplies. A refusal that names one is treated as what it is — a temporary limit — and the generation is retried automatically, the same way it is after any other passing interruption.
A genuine problem with the provider account still stops the work immediately, and that is deliberate: an exhausted budget or a rejected key cannot be waited out, and retrying would only delay the moment somebody is told about it.
Nothing about your own orders changes. Generations that were completing normally are unaffected, and no setting has to be adjusted.