What one image actually costs

Two ways the number you are quoted stops matching the number you are charged, both found by OpenRouter's own billing measurements on 2026-09-11, which we cite rather than repeat as our own:

• The quality flag can cost more than the model. Leaving quality unset on gpt-image-2 billed $0.006; setting it to highbilled $0.211 — 35× for one flag. It is the same mechanism behind the 20× gap in our own token table below: gpt-image-2 emits 200 output tokens per image at its default quality, gpt-image-1-mini emits 4,160.
• Published rates and billed rates drift apart. Riverflow V2.5 Pro lists $0.13 per image and billed $0.064. Where a measured figure exists for a row, the table above shows it beside the list price so you can see the gap instead of trusting either number alone.

Token-priced models only get a per-image figure when we have a hand-verified token count for that exact model at its default quality; the rest stay marked token-priced with no per-image number rather than carrying an invented one.

Every morning our monitor reads OpenRouter's public image catalog, the rate line on each kie.ai model page, and fal.ai's catalog API. Per-image rates pass through untouched. Megapixel rates reduce at one megapixel for a 1024×1024 image. Credit rates convert at the channel's own published dollar figure. Token rates convert only where we hold a hand-verified token count for that model. Every converted row says which conversion it used, and a channel that doesn't respond keeps its last verified row and date rather than being refreshed with a guess.

Cross-channel spreads only compare the same output size. Rows whose size we cannot normalise are shown but excluded from the comparison, because a spread between two different image sizes is not a price difference.

Prices are list prices and can change without notice — always confirm on the provider's page before you spend. We run no affiliate links with any channel on this page, so we have no financial stake in which one you pick.