AI Image Generator API Pricing (2026): Real Cost Per Image, Every Model

AI Image Generator API Pricing (2026): Real Cost Per Image, Every Model

DALL·E 3 cost four cents an image. That number is still the reference point most people carry around, and it's now both wrong and unavailable: OpenAI pulled DALL·E 2 and 3 out of the API on May 12, 2026. The cheapest hosted image from a brand-name lab today costs half a cent. Rent your own GPU and it costs a tenth of that again.

Which should make AI image generation API pricing easy to compare, and doesn't, because no two vendors sell the same thing. Google and OpenAI bill output tokens. Black Forest Labs bills megapixels. Stability and Recraft bill credits. Replicate bills GPU seconds. Ask what the cheapest AI image API is and the only defensible answer is a question back: how many images, at what quality, and can you wait?

Below is every price we could verify, the date we checked each one, and the volume math that decides which model actually wins your bill.

The cheapest AI image API is a volume question

There's no single winner. The conditional answer, by tier:

• Open weights on an aggregator: FLUX Schnell runs about $0.003 an image on Together or Fal, and the normalized floor at Runware sits near $0.0006. SDXL Turbo bottoms out around $0.0002.
• Your own GPU: SaladCloud benchmarked FLUX.1 Dev on a rented RTX 4090 at 992 images per dollar, or $0.00101 each, at a 99.78% success rate.
• Cheapest hosted flagship: OpenAI's GPT Image 2 and GPT Image 1 Mini both start at $0.005 an image on the low quality tier.
• Cheapest from Google: Nano Banana 2 Lite at $0.0336, halving to $0.0168 on the batch tier, which is the lowest price Google publishes.
• Native SVG: Recraft V4.1 Vector at $0.08. Nothing else in the field returns editable vector output at all.

That's a 400-fold spread between the top and bottom of this list, and at 100 images a month the whole spread is worth less than a sandwich. At 100,000 a month it decides whether you hire someone.

So find the cost cliff you're standing on first. Then argue about models.

Every model, every price, every date

The table below normalizes everything to one comparable unit, the cost of a single 1024x1024 image, which is the de facto convention because it's the only thing every billing model can be converted into. Two live traps to flag before you read it.

First, Google deprecated and shut down Imagen 4 Fast, Standard and Ultra on August 17, 2026, and now points migrations at Gemini 2.5 Flash Image. Any guide still quoting Imagen 4 Fast at $0.02 is quoting a model you can't call. Second, GPT Image 1 retires on October 23, 2026, so don't build on it now.

The other thing every table on the internet skips: these figures are output only. Edits, reference images and multi-image workflows add input-image tokens on top, which we'll cost later.

Midjourney deserves its one line and no more. As of 2026 it still ships through Discord and its web app, with no official public API, and every "Midjourney API" on sale is an unofficial wrapper driving a real account behind the scenes, ban risk included.

Before you paste any row of that into a budget deck, check its date column. A price without a date is a rumor.

Nobody agrees on what one image costs

Imagine a fuel station where one pump sells by the liter, one by the mile, one by the full tank and one by the minute you stand there. Every price is real. None of them compare until you convert. That's this market, and the conversion is the work.

Google and OpenAI bill tokens, then divide

Both vendors bill image output in tokens, and the per-image number you see quoted is a division sum someone already did for you. Nano Banana Pro charges $120 per million output tokens; a 1K or 2K image is 1,120 tokens, so $0.134, and a 4K image is 2,000 tokens, so $0.24. Nano Banana is simpler still: 1,290 tokens at $30 per million, which is $0.039. Google publishes the conversions directly, which is more than most vendors do.

OpenAI works the same way with more dials. GPT Image 1.5 bills image tokens at $8 in, $2 cached and $32 out per million, plus text at $5, $1.25 and $10. For a square image that lands at roughly $0.009 low, $0.034 medium and $0.133 high. Switch to portrait or landscape and high quality jumps to about $0.20, because you're buying more tokens for more pixels.

One consequence people miss: image-to-image and multi-reference edits push tokens into the input side too, and none of the headline tables cover that.

FLUX charges by the area

Black Forest Labs scales with megapixels. FLUX.2 Pro is $0.03 for the first megapixel and $0.015 for each one after; Klein 4B starts at $0.014 and adds $0.001 per extra megapixel. A 1024x1024 image is roughly one megapixel, so the base rate is your rate, until someone on your team asks for 4K and the invoice quietly triples.

Robin Rombach, who co-created Stable Diffusion and now runs Black Forest Labs, launched the FLUX API at 2.5 credits an image with FLUX 1.1 Pro at four, which at a cent a credit put the flagship at exactly DALL·E 3 money. He also calls FLUX one of the lowest per-image prices in production, and he would: treat vendor "cheapest" claims as direction, not measurement.

Credits are per-image pricing in a costume

Stability, Recraft and legacy FLUX all sell credits, and in each case a credit is a cent. Stability's Core is 3 credits, Ultra is 8. Recraft charges one credit for a raster image and two for a vector on its web plans. The abstraction adds nothing except a second conversion step and a balance you have to top up.

Replicate sells seconds, not pictures

Per-second GPU billing means your cost per image depends on how fast the model runs and whether the container was warm. Cold starts on unpopular Replicate models run 20 to 60 seconds, and practitioners comparing hosts put Replicate at 3 to 10 times the cost of Fal or Together for the same open-weight model at volume. Fal, by contrast, bills per image or megapixel with no markup on model output: Seedream V4 at $0.0333, Flux Kontext Pro at $0.0425, Nano Banana at about $0.0398.

The rule of thumb worth memorizing: aggregators beat going direct on open-weight models, and roughly tie on closed flagships.

At 100 images nothing matters. At 100,000 everything does.

This is the cost cliff: the point where the cheapest option stops being a preference and starts being a hiring decision. Below it, price is noise and you should pick on quality. Above it, the ranking flips more than once.

All figures are our arithmetic on the rates in the table above, output only, per month.

At 100 images a month, the gap between the cheapest row and the dearest is $13.38. Nobody should optimize that. Use the best model you can find, which for most people means the Google or OpenAI flagship, and spend the saved afternoon on prompts instead.

At 1,000 a month you're still under $150 at the top of the table, which is one seat of most design software. Quality still wins.

At 10,000 the shape changes. The difference between Nano Banana Pro and Nano Banana 2 Lite on batch is $1,172 a month, which is real budget, and the question becomes whether your users can tell. Run both, look at fifty outputs side by side, and decide with your eyes.

At 100,000 you're in a different business. Nano Banana Pro at 4K costs $24,000 a month. The same volume of GPT Image 1 Mini low-tier images costs $500. Self-hosted FLUX.1 Dev, compute only, costs $101. The self-host row is where most people's arithmetic goes wrong, so it gets its own section below.

Bonus info: if you're budgeting in dollars per month rather than dollars per image, we ran that version of the math separately in How Much Does AI Image Generation Cost?.

Batch halves your bill if you can wait

Google and OpenAI both run a batch tier at 50% off, and it's the most under-used line item in this whole market. Nano Banana Pro drops from $0.24 to $0.12 at 4K. Nano Banana goes from $0.039 to $0.0195. Nano Banana 2 Lite lands at $0.0168, the cheapest thing Google sells. OpenAI applies the same 50% to its Batch API, and discounts cached input on top.

The trade is asynchronous delivery. If a human is watching a spinner, batch is useless. If the images feed a product catalog, a scheduled newsletter, a thumbnail pipeline or an overnight render queue, you're leaving half the invoice on the table by not using it.

Two smaller levers worth knowing. OpenAI gives new accounts $5 in free credits, which is around 1,000 GPT Image 1 Mini images at the low tier: enough to run a real evaluation before you commit. And Stability's 25 free credits are one-off, with no recurring free tier, so don't plan around them.

Sort your image generation into "someone is waiting" and "nobody is waiting" this week. Move the second pile to batch.

Self-hosting costs $0.001 an image and one engineer

The compute floor is real and it's been measured. SaladCloud ran FLUX.1 Dev on RTX 4090s at batch priority and recorded 992 images per dollar, which is $0.00101 an image, at 20 steps and 1024x1024, with sub-18-second average response times and a 99.78% success rate. That's five times cheaper than the cheapest hosted flagship and roughly 130 times cheaper than Nano Banana Pro.

Now the arithmetic nobody puts next to it. A rented 4090 runs $0.34 to $0.69 an hour, so holding one for a month at fifty cents costs about $360, and $360 buys 72,000 GPT Image 1 Mini images at the hosted rate. Below roughly 70,000 images a month, the API is cheaper than the GPU before anyone has written a line of code. Above it, the GPU wins on compute and starts losing on everything else: cold starts, idle capacity, a pipeline someone has to keep alive, and a model refresh every two to four months. Analyst estimates put a production self-hosted pipeline in the hundreds of thousands of dollars a year in engineering time, and those come from calculators with an axe to grind in the other direction, so read them as directional.

Think of it like buying the office espresso machine. Cheap per cup, absolutely, and someone still has to descale it every Friday.

Self-hosting wins on sustained, predictable, high volume with a GPU you can keep busy. It loses on spiky traffic, on small teams, and on anything where the person maintaining it has better things to build. Both benchmark figures here come from GPU vendors with an interest in making rented silicon look cheap, which is why the break-even matters more than the floor.

The costs that never make it into the table

Every price above is a base fare. Here's the catch: the extras are where budgets break, in the same way an airline seat costs $49 until the bag, the seat and the priority lane.

• Input tokens on edits. Image-to-image and multi-reference workflows bill the images you send in as well as the one you get back. Published per-image tables cover output only.
• Retries and rejections. Most vendors don't charge for failed generations, though some wrappers do, and moderation refusals still cost you the round trip and the user's patience.
• Upscaling. Recraft's Creative Upscale is 20 credits, twenty times a raster generation, which reframes "cheap per image" fast.
• Licensing. FLUX.2 Dev is non-commercial without a license. Recraft's free-plan outputs are public and not licensed for commercial use. Stability's Community Licence is free for commercial use under $1M in revenue.
• Storage, egress and floors. Images are heavy, CDN bills are real, rate limits cap throughput per tier, and some platforms impose a minimum balance, such as the ¥200 floor on Volcano Engine.

Character references add a multiplier too: Ideogram's character-reference feature costs 1.7 to 3.3 times the base rate, so a $0.03 Turbo image can become a $0.10 one without changing model.

Take your last month's generation log, tag each call as new, edit, upscale or retry, and price the four buckets separately. The number you get is the one your CFO will see.

Five API keys, or one subscription

Now the part where we're conflicted, so let's put it on the table. JammyJar routes these models under a single subscription, which means we have an obvious interest in you concluding that stitching APIs together is a hassle. Judge it on the arithmetic below, and treat our margin as another cost in the ledger.

The case for going direct is strong at scale. You pay the model price, nothing more, and at 100,000 images a month that difference dwarfs any convenience fee. The case against is the maintenance: five keys, five billing relationships, five rate-limit ceilings, and five deprecation calendars. In the last four months alone that calendar removed DALL·E from the API, shut down Imagen 4 and put an October date on GPT Image 1.

Subscriptions can also beat APIs outright, which is the finding vendors bury. Recraft's Basic plan is around $10 a month for 1,000 credits, roughly a cent per raster image against $0.035 on the API. Analysts pricing Ideogram found its app subscription undercuts its own API by about five times at high volume. If you generate steadily and never need programmatic access, the subscription is frequently the cheaper product.

So the rule is boring and it holds: below a few thousand images a month, optimize for the number of things you have to maintain. Above it, optimize for the per-image rate, and go direct.

If you'd rather compare models than compare invoices, JammyJar gives you one key across these models with PNG, SVG, WebP and JPG export, and no manuals.

AI image generation API pricing questions, answered

What's the cheapest AI image generation API? For hosted models, GPT Image 2 and GPT Image 1 Mini start at $0.005 an image on the low quality tier, and Google's Nano Banana 2 Lite drops to $0.0168 on batch. Open-weight models on aggregators go lower: FLUX Schnell runs about $0.003, and Runware's normalized floor sits near $0.0006 an image.

Does Midjourney have an API? No. As of 2026 Midjourney still ships through Discord and its web app only, with no official public API. Every product advertising a Midjourney API is an unofficial wrapper driving a real account behind the scenes, which carries an account-ban risk. For programmatic generation, pick a model with a supported API instead.

Is Imagen 4 still available? No. Google deprecated and shut down Imagen 4 Fast, Standard and Ultra on August 17, 2026, and directs migrations to Gemini 2.5 Flash Image, sold as Nano Banana. Any pricing guide still listing Imagen 4 Fast at $0.02 an image is stale. Nano Banana costs $0.039, or $0.0195 on the batch tier.

How much does GPT Image cost per image? It depends on quality and resolution, because OpenAI bills image output in tokens. GPT Image 2 spans $0.005 to $0.211. GPT Image 1.5 works out at roughly $0.009 low, $0.034 medium and $0.133 high for a square image, and about $0.20 for a high-quality portrait or landscape. Edits cost extra as input tokens.

Which AI image API returns real SVG? Recraft, and effectively only Recraft. Recraft V4.1 Vector returns editable SVG at $0.08 an image, with V4.1 Pro Vector at $0.30. Every other major model returns raster output you'd have to trace afterwards. For logos, icons and anything that must scale cleanly, that makes Recraft a category of one.

Do batch APIs really halve the price? Yes, on both Google and OpenAI, in exchange for asynchronous delivery. Batch takes Nano Banana Pro's 4K image from $0.24 to $0.12, and Nano Banana 2 Lite from $0.0336 to $0.0168. If your images feed a catalog, a newsletter or an overnight queue rather than a live user, batch is the easiest 50% you'll ever save.

Where these numbers came from

Every price came from the provider's own pricing page where one exists, checked on August 19, 2026, except where the table says otherwise: OpenAI's ranges were checked August 2, Stability in July, and Ideogram across May and June. For token-billed models, the per-image figures are the conversions the vendors publish themselves, not our estimates. ByteDance and Alibaba rates are aggregator-derived and marked approximate, because their official consoles are token-based and gated to the Chinese market, and we couldn't verify them independently from outside China. Monthly totals are our arithmetic on those rates, output only. Third-party resellers advertising 70% or 80% discounts are excluded on purpose: they're reselling the same models as loss leaders or under terms that can change without warning. We re-check quarterly, and prices in this market have halved on new tiers within a single quarter, so anything here older than three months is indicative rather than quotable.

Wire up one model this week

Go back to the four cents. DALL·E 3's price was easy to remember because there was one number and one model, and that world closed on May 12. Now there are five billing units, four quality tiers per vendor, a batch lane that nobody uses, and a shutdown notice every few months. Cheapest isn't a model. It's a row in the volume table, and the cost cliff decides which row you're in.

So find your row. Take the two cheapest models in it, run the same fifty prompts through both, and price the output against what you'd have paid on your current setup. Your bill will tell you more in a week than any table will, including this one.

Or skip the five keys and let JammyJar route them for you.