The saving with no quality trade-off
Most providers discount asynchronous batch processing by around half, in exchange for a latency budget measured in hours rather than seconds. The model and the output are identical.
It is the only lever on this site that costs nothing in quality, and it is skipped mostly because it needs a slightly different code path.
What can move
Nightly classification, backfills, evaluation runs, embedding generation, enrichment pipelines, summarising yesterday’s tickets — anything where no user is waiting on the response.
Teams usually find the batchable share is far higher than they assumed, because interactive traffic is a smaller fraction of total requests than it feels like.