How we cut image generation costs by 38%
Last quarter our image bill grew faster than our user base. Most of the spend went on drafts that people looked at once and threw away, so we changed where those drafts are made.
Drafts now go to a fast, low-cost model. Only the draft a person picks is rendered again on the large model. Because most sessions end after one pick, large-model calls fell by 61%.
We also cache results. When the same prompt and settings come in within 24 hours, we return the stored image instead of generating a new one. About 9% of requests hit the cache, mostly from shared templates.
Finally, free accounts now get 20 images a day, down from 50. Paid plans are unchanged.
Quality held: the average rating stayed at 4.4 out of 5 across the quarter.