Cost math for AI images that survive a finance review

Start with dated, reproducible inputs

The seed values on this page were supplied by the operator for one 960×1472 portrait: $0.0006 on Runware schnell, approximately $0.0037 on Together, $0.0045 on Leonardo, and $0.007 on Runware dev. They are internal measurements, not guaranteed public prices.

Re-run them from the production account before publishing. Record model ID, provider, endpoint version, width, height, steps, output count, region, sample size, success rate, date, and whether discounts or reserved capacity apply. A comparison without configuration is not evidence.

Calculate the fallback-weighted provider cost

If 94% of fast requests succeed on Runware, 4% move to Together, and 2% move to Leonardo, the provider component is a weighted sum. Include the failed primary attempts when they incur a charge; the simple calculator assumes only the successful route is billed until you enter better data.

Cost is often roughly sensitive to pixel area, but provider billing models vary. The page scales the seed measurement by requested pixels only as a planning approximation. The server returns actual provider-reported cost when available and labels other values as estimates.

blended provider cost = Σ(route share × route cost)

cost per accepted image =
  (provider + moderation + storage + retry overhead) / acceptance rate

Set customer credits from the conservative case

Add payment processing, support, observability, fraud, taxes where applicable, and gross-margin target. Then stress the model: fallback share doubles, premium use rises, moderation gets more expensive, or acceptance falls. A plan that only works in the best week is not a plan.

A simple launch structure can assign one credit to a fast image and several credits to premium, while keeping the dollar value of a credit comfortably above the worst eligible fast fallback. The exact price is a business decision and should be validated against customer value, not only compute cost.

• Use p95 blended cost, not only the mean
• Cap batch size and concurrency by plan
• Alert on cost per accepted image
• Review provider prices and routes monthly
• Do not advertise permanent savings from one benchmark

Publish benchmark methods buyers can inspect

A dated methodology page can earn more trust than dozens of generic ‘best API’ articles. Publish the sample prompts or prompt categories, dimensions, model identifiers, run count, error handling, and calculation workbook. Explain limitations and update history.

Do not manufacture pages for every trivial cost variation. Keep one canonical calculator, one detailed benchmark report, and comparison pages that answer different buyer decisions. This creates depth without scaled-content spam.