What does an AI-generated image actually cost?

"How much does an AI image cost?" ought to be a simple question. It is not, because image models do not bill per image. They bill per token — the same unit as text — and an image is worth some number of tokens that depends on its resolution.

So the pricing page gives you a rate per million tokens, and you are left to work out how many tokens your picture is.

Only one vendor does the arithmetic for you

This is the practical finding. As of 17 July 2026:

Google publishes per-image equivalents right on its pricing page:

OpenAI publishes token rates only, and directs you to a calculator in its image generation guide:

We are not going to invent a per-image figure for OpenAI. Converting a token rate to a picture price requires knowing the tokens-per-image at each resolution and quality setting, and OpenAI does not publish a single headline number — that is what the calculator is for. Any blog post quoting you a confident cents-per-image for gpt-image-2 either used the calculator for one specific configuration, or made it up.

The caveat that spoils even Google's table

Look closely at that first table. The three per-image figures are quoted at different resolutions: 1K, 0.5K, and 1K/2K respectively.

That means they are not comparable to each other, let alone to anything else. A half-resolution image from the Flash model costing more than a 1K image from the Lite model is not a contradiction — they are different pictures.

The token rate is the only honest common axis, and even it moves with resolution, because resolution is what determines the token count.

What you can actually conclude

• The rate spread is roughly 4× between the cheapest and most expensive image output on this page ($8/1M to $32/1M), and 4× again across Google's line ($30/1M to $120/1M).
• Image tokens are expensive relative to text. Google's most expensive text output is $18/1M; its image output runs to $120/1M. Pictures are the premium product, not a rounding error on your bill.
• Caching applies to image input — unevenly. OpenAI discounts cached image input by 75% on gpt-image-2 and gpt-image-1.5 ($8.00 → $2.00), but by 90% on gpt-image-1-mini ($2.50 → $0.25). The cheapest model gets the best discount, which is the opposite of what you would guess. Worth doing if you re-send the same reference image — though at 75% it saves you about as much as dropping a model tier, not more.
• Cents per picture. Google's published range is roughly 3 to 13 cents. A thousand images costs $34 to $134, depending on model and resolution.

How to actually budget this

1. Pick your resolution first. It is the variable that determines cost. Every published figure is conditional on it.
2. Use the vendor's own calculator for OpenAI, and its per-image equivalents for Google. Do not trust a third party's cents-per-image — including anyone quoting one without stating a resolution.
3. Generate ten at your real settings and read the bill. This is the only method that produces a number you can plan with, and it takes ten minutes.
4. Cache your reference images if you send the same ones repeatedly.

The uncomfortable summary: the honest answer to "what does an image cost" is "measure it". Everyone quoting you a clean number is quoting a configuration, usually not yours.