Image APIs are priced per generated image, but you rarely keep every image you pay for. Most teams generate several variations and pick one, and some requests produce nothing usable at all. This calculator turns the price on the pricing page into the number that matters for a budget: the cost of each image you actually use.
How it works
Start with the images you need. If 20% of requests produce nothing usable, you need 1 ÷ (1 − 0.20) = 1.25 requests for each keeper. Multiply by the variations per request to get the images you pay for.
Monthly cost is images generated × price per image. Cost per usable image divides that by the images you keep. The ratio between the two is your real overhead: generating 4 variations with a 20% failure rate makes each keeper cost 5 times the sticker price.
Storage is small per image but it accumulates. Each month adds a new batch, so after a year you pay for twelve months of images at once. Storing only the images you keep, and deleting rejected ones, keeps this in check.
Prices differ by model, resolution and quality setting, and change often. Enter the current price from your provider's pricing page for the exact settings you use. The default is an example, not a quote.
A worked example
You need 2,000 product images a month and generate 4 variations per request at $0.04 each. One in five requests is unusable, so you make 2,500 requests and 10,000 images, costing $400.00 a month. Each image you keep really costs $0.20, five times the sticker price. Keeping 2 MB per kept image adds 4 GB a month.
Questions people ask
How much does AI image generation cost per image?
Prices range from fractions of a cent to several cents or more per image, depending on the model, resolution and quality. The sticker price is only part of it: variations and failed attempts usually multiply the real cost per usable image several times over.
How can I reduce AI image generation costs?
Generate fewer variations once your prompts are reliable, draft at a lower resolution or quality and upscale only the winners, reuse prompt templates that already work, and track your reject rate so you can see which prompts waste money.
Is it cheaper to self-host an image model?
At high, steady volume it can be, but you pay for GPUs whether or not they're busy, plus engineering time to run them. Compare the monthly API cost from this calculator with the full monthly cost of the GPU servers you'd need at peak load.
Why count variations and rejects separately?
They have different fixes. Variations are a choice you control per request. Rejects come from prompt quality or model limits, like text in images or hands. Measuring both shows where the money actually goes.
Do I own the images an AI API generates?
It depends on the provider's terms and your local law. Most commercial APIs let you use the outputs, but read the current terms of service, especially for trademarks, real people and content for clients.