AI image generation cost calculator

How to calculate what a batch of AI images will cost

Three inputs, all on this page. The calculator does not call any API and does not need a key — the prices are constants published on this page with their sources.

1. Count the images you need to keep, not the calls. Budgets are written in finished assets: 40 product shots, 12 hero images, one avatar per team member. Put that number in the first box.
2. Put in your real throw-away rate. Almost nobody keeps the first generation. If you discard 3 of every 10, enter 30 — the calculator runs the batch to 58 generations, because the re-runs get thrown away at the same rate too.
3. Read the batch column, not the unit price. The table sorts by what the whole batch costs. Per-image prices that differ by a tenth of a cent stop mattering the moment your retry rate moves five points, and that is the point of the comparison.

What a batch of AI images actually costs

A generated image costs between $0.003 and $0.09 at published API prices, and about $0.05 on Clize's hosted pricing — but the number you are actually going to pay is the per-image price multiplied by every generation you throw away. At a 30% throw-away rate, 40 keepers take 58 generations, so a $0.04 model bills $2.32, not $1.60. That single multiplier moves the bill more than the choice of vendor does.

So the honest form of the question "how much does AI image generation cost?" has three inputs, and the calculator above asks for all three: how many finished images you need, how many generations it takes to get one you keep, and how many reference images you feed each call. Resolution and aspect ratio matter on token-priced APIs and not at all on flat per-image ones; that difference is in the table below.

Two things the calculator deliberately does not do. It does not call anyone's API — the prices are constants on this page, each with the vendor page it came from and the date it was read, so you can check them rather than trust them. And it does not quote a price for a model whose vendor has not published one: OpenAI's gpt-image-2 is billed per token, and any per-image figure for it is somebody's arithmetic, not a price.

Published per-image prices, with sources

Every row is the vendor's or marketplace's own published figure, read on 4 September 2026. The calculator above uses exactly this table — nothing is estimated, averaged or marked up in between.

Prices change, and a stale table is worse than no table. Every row carries the date it was read, the calculator repeats those dates in its output, and this page's visible Updated line moves whenever the table does. If you find a row that has drifted, the vendor's link is right there — believe it over us.

The number that moves the bill: retries

Every cost calculator you will find asks for a count and a unit price and multiplies them. That is not how image generation is bought. You do not buy images, you buy usable images, and the gap between the two is where the money goes.

The arithmetic is a keep rate, not a markup. If you throw away a share r of everything you generate, then to end up with K keepers you must run K / (1 − r) generations — because the re-runs get thrown away at the same rate. At a 50% throw-away rate that is twice the calls, not 1.5×. Rough anchors from real batches:

• 10–20% — a prompt you have already tuned, a familiar style, no text in the image.
• 30–40% — the normal case: a new brief, a house style you are still pinning down.
• 50–70% — legible text in the image, a specific human face, an exact product in an exact pose. Reference images help here more than a bigger budget does.

Which is why the practical move, when a batch is over budget, is almost never "switch to the model that is a cent cheaper". It is to cut the throw-away rate: tighten the prompt, send reference images instead of describing the thing in words, and generate one image before you generate forty. Sending references is free on Clize, so that lever costs nothing to pull.

Reference images: what the caps are, and what they add to the bill

Reference images — a product photo, three views of a character, a colour board — are the difference between describing a thing and showing it. On Clize's hosted generation they add nothing to the price: an image is $0.05 whether you send none or sixteen. What they do have is a cap, set by the model:

There is a size ceiling too: any one reference image must stay under 10MB once encoded, and the whole set under 80MB, because they travel in a single request. Over the cap the call is refused before anything is charged, which is the behaviour you want — a rejected request costs nothing.

On APIs billed per token the arithmetic is different and worth knowing: the reference images you send are input, and input is metered. On OpenAI's gpt-image-2 that is a separate image-input rate on top of the output tokens, which is one more reason no honest per-image number exists for it. On the flat per-image marketplaces in the table, an image-to-image call is priced like any other call.

One flat price, and a quote before anything is charged

Clize's hosted generation is $0.05 an image across every model it offers, every size, with or without reference images — and $0.48 for a video clip. That is the whole price list. The reason it is worth a flat rate rather than a fractionally cheaper per-call price is that a flat rate is the only one an agent can forecast: forty images is $2.00 before you have written the prompt, and it is still $2.00 after you switch models mid-batch.

The money gate is the part that matters when the thing spending is an agent rather than you. Every generation is quoted before it is charged: clize gen image "…" on its own prints the price and spends nothing, and only adding --confirm actually generates and charges. There is no pre-approved allowance and no way to skip that flag — each generation is confirmed on its own, every time. If a generation fails or the provider refuses it, the charge is refunded in full, including a video job that fails after it was submitted.

$ clize gen image "product on a white cyclorama, soft key light"
      📋 Quote about $0.05 (image). Add --confirm to generate.
      
      $ clize gen image "product on a white cyclorama, soft key light" --confirm
      ✓ clize-assets/gen_1a2b_0.png

Everything already spent is on the record: clize gen list shows every asset with its model, prompt and cost, and clize status shows the month's total. Generation is one of several things Clize does with money in the loop — the same quote-then-confirm gate covers domain registration and the storefront. If you are pricing video instead, the video cost calculator runs the same comparison on per-second pricing, and Agent Media is the product these prices belong to.

$ npm i -g @clize/clize && clize install
$ clize gen image "a red cube on white" --ref board.png