How much does AI image generation cost per image?

As of September 2026, API generation runs from under a cent to about twenty cents per image depending on model, size, and quality tier: Flux 2 from about $0.015, Nano Banana 2 from $0.045 at 0.5K to $0.151 at 4K, GPT Image priced per token with third-party trackers estimating a few cents to around eight cents for a standard image. Subscriptions replace the per-image price with quotas: ChatGPT Plus $20, Midjourney $10 to $120, Canva Pro $18 with 500 generations.

What is cost per accepted image?

Total generation spend divided by the number of images you actually published. It is the only figure that reflects reality: a three-cent image that takes four tries costs twelve cents, plus the time. Batch-of-four previews and automatic quality tiers inflate it without appearing on any pricing page.

Is a cheaper AI image model actually cheaper for a marketing team?

Often not. The regeneration rate dominates the bill, and regeneration is driven by how on-brand the first output is, which depends on the brand profile and reference images attached, not the model's price. Consistency tooling that cuts tries from three to one saves more than switching to a model that costs a third as much.

How much should a small marketing team budget for AI images?

For a realistic month of about 55 images (social tiles, email heroes, blog covers, landing pages), raw generation at API prices is a few dollars to a few tens of dollars even with regenerations. The real budget line is the tool or subscription that makes those images on-brand and puts them where they go; that is where the difference between $20 and $200 a month lives.

← All posts

What AI Image Generation Actually Costs a Marketing Team: Cost per Accepted Image, Not per Generation

Every pricing page in this category quotes a price per image, and every comparison article dutifully copies it into a table. Those numbers are true and they are the wrong numbers. Nobody ships every image they generate. The image that costs three cents to make and takes four tries to get on brand cost twelve cents and fifteen minutes, and the pricing page has no row for that.

This post prices AI images the way a marketing team actually pays for them. The three pricing models and what they hide, a realistic month for a small team, the two multipliers that quietly dominate the bill, why the cheapest model is rarely the cheapest option, and a worksheet. Prices are as of September 2026 and linked; every vendor here has changed its pricing within the year.

The three pricing models

Seats. ChatGPT Plus at $20 a month, Midjourney from $10 Basic to $120 Mega, Canva Pro at $18. You do not pay per image; you pay for a quota. ChatGPT Plus is tracked at roughly 50 images per three-hour window; Midjourney meters fast GPU hours by tier; Canva Pro includes 500 Magic Media generations a month. The seat is cheap if you use the quota and expensive if you do not, and it says nothing about quality per image because there is no lever to pull.

Credits. Freepik, Leonardo, Recraft, and most multi-model aggregators sell credit packs, with different models consuming different credits. Readable once you learn the exchange rate, opaque until then, and the mechanism by which “a $10 plan” becomes “$10 for 200 images or 40, depending.”

Tokens or per-image API pricing. The transparent model, and the one that tells you what an image actually costs. Google publishes Nano Banana 2 per image by resolution: $0.045 at 0.5K up to $0.151 at 4K. Flux 2 runs about $0.015 to $0.05. OpenAI bills GPT Image per token, $30 per million output tokens on the 2.5 models, with output tokens rising by quality tier; third-party trackers have put a standard GPT Image 2 image at a few cents to around eight cents, and OpenAI has said its GPT Image 2 calculator does not apply to 2.5. Batch APIs from OpenAI and Google cut roughly 50% for jobs that can wait.

The whole market sits between roughly half a cent and twenty cents a generation. Which is to say: the raw generation is nearly free, and everything that costs money is what happens around it.

A realistic month

A small team, one marketer plus a founder who reviews, shipping what such teams ship:

Asset	Count per month	Typical tier	Notes
Social tiles (LinkedIn, Instagram)	32	medium	Two per business day; illustrated
Email heroes	8	medium	Weekly newsletter plus nurture sends
Blog covers	4	medium	One per post
Landing page heroes	2	high	Edited several times each
Campaign hero (the one everyone has opinions about)	1	high	Edited five or six times
Ad and one-off variants	8	medium	Seasonal offer, event tile
Accepted images	55

At three to eight cents per generation, 55 accepted images is somewhere between $2 and $5 if every image were accepted on the first try. It never is. Now the multipliers.

The two hidden multipliers

Multiplier 1: the regeneration rate

How many generations per accepted image. Without a brand profile and reference images, teams we have watched land between two and four tries for a social tile and six to ten for the campaign hero; the model reaches for its default look, someone rewrites the prompt, repeat. With a profile and references attached to every generation, the tile drops to one or two and the hero to two or three, because the first output is already in the right palette and style and the edits are local. Same model, same price per generation, a two- to three-times difference in the bill, and a much larger difference in the afternoon.

Multiplier 2: images per request, and who picks the quality

Two product decisions, invisible on pricing pages, decide the rest.

Many tools generate four images per request so you can pick one. That is four generations billed for one preview most people glance at, and it quadruples the cost of every try before regeneration is even counted. One image per turn, then iterate, costs a quarter as much and produces a better image, because iteration converges and a grid of four does not.

Quality tiers are the second decision. GPT Image offers six (low to max on the 2.5 models); the auto setting lets the model choose. In our production runs on GPT Image 2, moving one tier from medium to high multiplied cost per image by roughly five, with a visible improvement on a print-sized hero and none on a phone-sized social tile. A tool that runs auto is letting the model spend your money whenever the prompt sounds important, and the model finds most prompts important.

Put both multipliers together and the same 55 images cost anywhere from a few dollars (one image per turn, tier pinned per asset, profile and references attached) to well over a hundred (four per request, auto quality, prompts from scratch). Nothing on any pricing page distinguishes the two.

The number to track. Cost per accepted image: total spend on generation divided by images published. Track it monthly. If it is falling, your brand profile and references are working. If it is flat while your model bill falls, you switched to a cheaper model and are regenerating more, which is the most common way to save nothing.

Why consistency tooling beats a cheaper model

Suppose you are choosing between a model at three cents and one at eight, and the eight-cent model holds your reference subject through edits while the three-cent one does not. For the campaign hero, six tries on the cheap model is eighteen cents; three on the expensive one is twenty-four. Close. Now add the fifteen minutes per try, and the cheap model cost you 45 extra minutes to save six cents. Now add that the cheap model’s six outputs disagree with each other and with last week’s set, so one of them gets published anyway and reads as off-brand.

The lever that lowers cost per accepted image is not model price. It is:

A brand profile on every generation (palette, imagery kind, never-list), so the first output is in range. Turning Brand Guidelines into an AI Image Brief is the template.
Reference images attached, which in our GPT Image 2 runs added roughly 40% to per-image input cost and paid for itself in the first avoided regeneration. The reference-images guide explains the set.
Library first, so an image you already have is reused instead of regenerated. Free.
One image per turn, quality pinned per asset type, so a prompt cannot escalate the bill.
Local edits instead of reprompts, which GPT Image 2.5’s comments and edit stability made much cheaper.

Model choice still matters (run the six tests), but it is the smaller lever, and Flare vs Sunburst is the case study: same price per token, and the slower model is cheaper on the hero because it needs fewer tries.

The worksheet

Fill it in for last month. It takes ten minutes and it will change what you buy.

Images published last month:                        ____  (A)
Generations run (or tries, if a subscription):      ____  (B)
Regeneration rate:                                   B / A = ____
Images per request (1 or 4):                        ____
Quality tier: pinned per asset / auto:              ____
Spend on generation (API) or subscription cost:     $____ (C)
Cost per accepted image:                             C / A = $____
Minutes per accepted image (honest estimate):       ____
Images regenerated because off-brand (of B):        ____

If the regeneration rate is above two, the fix is a profile and references, not a model. If images per request is four, the fix is a tool that generates one. If the tier is auto, pin it. If minutes per accepted image is above ten, the cost is not the generation at all; it is the download-rename-upload loop, which The Hidden Cost of Copy-Paste Marketing prices separately.

What we pin

Since this post asks you to interrogate every tool’s defaults, here are Marqeable’s: one image per turn, then iterate; quality tier pinned in code per asset type, so neither a prompt nor the model can escalate it; the brand profile and a rotating set of reference images attached to every generation; your library searched before anything is generated; and the image created inside the email, post, or page it belongs to, so the minutes-per-image column drops with the cents. Generation runs on OpenAI’s GPT Image 2.5 models. It is not the cheapest model on any pricing page; it is chosen to have the lowest cost per accepted image for the assets our users make. We are in private beta with a small early cohort: get early access if you would like to see the worksheet filled in on your own month.

Frequently asked questions

Is ChatGPT Plus the cheapest way to get good AI images?

For an individual generating a few images a day, yes: $20 for a top-tier model with reference support. For a team, it becomes the most expensive option per accepted image, because it has no brand memory between chats and every image is re-prompted from scratch.

Is Midjourney worth $30 or $60 a month for a marketing team?

For concepting and hero art, a $10 Basic seat is worth it. For the monthly set, the metered GPU hours and the lack of an API or brand memory push cost per accepted image up. Midjourney Alternatives for Marketers covers where it should stay.

Do batch APIs make sense for marketing?

For anything that can wait hours (a month of blog covers, seasonal variants), yes, roughly half price. Not for the iterative work, which is most of it.

Should I pay for high quality?

For print and landing-page heroes, yes. For anything viewed on a phone for two seconds, no; the difference is invisible and the cost is several times higher. Pin it per asset type.

The bottom line

The raw generation is nearly free and the pricing tables are honest about that. What costs money is regenerating until the image is on brand, generating four to pick one, and letting the model choose the quality tier, none of which appears on a pricing page. Track cost per accepted image. Lower it with a brand profile, reference images, library-first sourcing, one image per turn, and pinned tiers. Then, and only then, argue about which model is cheaper.

Marqeable runs your campaigns, answers every visitor, text, and email in seconds, and turns them into booked jobs and meetings - even at 9pm on a Saturday. We’re in private beta with a small early cohort. Get early access