How do you calculate the real cost of a generated image?

Divide the list price per generation by your acceptance rate. At one approved image in eight, an asset listed at a few cents actually costs closer to a quarter.

What lowers image generation cost the most?

Anything that raises the acceptance rate: iterating at low resolution before upscaling the chosen result, a prompt library for known styles, reference images and fixed seeds, and a written brief before the first generation.

Why is video generation harder to budget?

A clip can fail on motion, temporal consistency or a single bad frame, and you pay for the full clip regardless. Per-second pricing makes iteration materially expensive very quickly.

Quick answer: Image and video pricing looks unusually simple: a price per image, or per second of output. That simplicity is why almost every team budgets it wrong. The provider charges per generation. Your...

Cost per approved asset in image and video

Updated September 1, 2026 · first published September 1, 2026

Image and video pricing looks unusually simple: a price per image, or per second of output. That simplicity is why almost every team budgets it wrong. The provider charges per generation. Your business consumes approved assets. Those are not the same number, and on visual work they are usually not even close.

The acceptance rate is the whole model

Nobody ships the first render. A designer generates, looks, adjusts the prompt, generates again. For a well-defined style with a mature prompt, acceptance might be one in three. For open-ended creative exploration, one in ten or worse. Video is harsher still, because a clip can fail on motion, on temporal consistency, or on a single bad frame — and you pay for the whole clip either way.

So the useful arithmetic is not the sticker price. It is sticker price divided by acceptance rate. At one in eight, an asset with a list price of a few cents costs closer to a quarter, and a video generation priced per second becomes a genuinely material line the moment a team starts iterating.

What actually moves the number

Resolution and duration are the obvious levers and the least interesting ones, because they trade directly against the output you wanted. The levers that matter are the ones that raise acceptance. Iterating at low resolution and only upscaling the chosen frame changes the economics more than any per-image discount. A prompt library for known styles turns one-in-ten exploration into one-in-three execution. Reference images and seeds convert a fresh gamble into a controlled variation. And a written brief before the first generation removes the most expensive failure mode there is: iterating toward a target nobody had defined.

Measure the right thing

Track generations per approved asset, per project and per person, and cost per approved asset as your headline number. Both are cheap to instrument and immediately actionable — a team at fifteen generations per approval has a brief problem or a tooling problem, not a pricing problem, and no provider switch will fix it.

Want this applied to your own LLM spend? FinOps LLM runs a free audit of your AI costs and shows where the savings are. Book free audit →

Back to research