Image generation APIs priced per 1,000 images, checked on official pages on October 6, 2026. Compare the $3, $25 and $90 tiers at 10,000 and 100,000 images a month, see how rejected generations change the real cost, and find out when a flat-fee API wins.
The cheapest image generation API for developers costs about $3 per 1,000 images, three tenths of a cent per picture. The most expensive popular option on the same price sheets charges $150 for the same 1,000. That fifty-fold gap decides whether an image feature is a rounding error or your biggest line item. This article lines up prices checked on official pages on October 6, 2026 from Replicate, fal.ai and Pruna, converts every one to a cost per 1,000 images, and shows what each choice does to a monthly bill at 10,000 and 100,000 images. It also breaks down the hidden costs that quietly erase a low sticker price, and what the PicassoIA API offers developers who want a flat fee instead of a meter.
Providers quote prices in three different units. Replicate lists some models per output image and others per thousand images. fal.ai bills several models per megapixel. Pruna advertises one flat price per image. Comparing them is guesswork until you convert everything to one unit, so every figure below is expressed as cost per 1,000 images at about one megapixel, the size of a 1024 by 1024 picture.
Each number in this table comes from the provider's own pricing or model page. Where a provider bills per megapixel, a 1024 by 1024 image is 1.05 megapixels, so the per-1,000 figure is marked as approximate.
| Model | Provider and billing unit | Listed price | Cost per 1,000 images |
|---|---|---|---|
| FLUX Schnell | Replicate, per thousand images | $3.00 | $3.00 |
| FLUX Schnell | fal.ai, per megapixel | $0.003 | about $3.00 |
| P-Image | Pruna, per image | $0.005 | $5.00 |
| Z Image Turbo | fal.ai, per megapixel | $0.005 | about $5.00 |
| FLUX Dev | Replicate, per image | $0.025 | $25.00 |
| Nano Banana | fal.ai, per image | $0.0398 | $39.80 |
| FLUX 1.1 Pro | Replicate, per image | $0.04 | $40.00 |
| Recraft V3 | Replicate, per image | $0.04 | $40.00 |
| Nano Banana 2 | fal.ai, per image | $0.08 | $80.00 |
| Ideogram v3 Quality | Replicate, per image | $0.09 | $90.00 |
| Nano Banana Pro | fal.ai, per image | $0.15 | $150.00 |
💡 Prices move. The sources were replicate.com/pricing, fal.ai/pricing and Pruna's own product page. OpenAI and Stability AI price by quality tier and credit bundle, so they are not in the table. Apply the same conversion to their pricing pages before you sign anything.
The two billing styles behave very differently once you leave 1024 by 1024. A megapixel meter charges for pixels, so a 2048 by 2048 image (4.19 megapixels) costs about four times a 1024 by 1024 one. At $0.003 per megapixel, that is roughly $0.0126 per image, or $12.60 per 1,000. A flat per-image price usually holds across a range of sizes, but check the model page for the exact limits.
Megapixel billing is cheaper for small thumbnails and more expensive for large hero images. If your product mostly needs 512 by 512 previews, the meter works in your favor. If you need print-size output, the flat price is safer to budget.
Three models sit in the bottom price bracket, between $3 and $5 per 1,000 images. They share a trait: each one is built for speed first and maximum fidelity second.
FLUX Schnell is the price floor in every list checked for this article. Replicate lists it at $3.00 per thousand output images, and fal.ai's $0.003 per megapixel lands in the same place at one megapixel. It is a distilled model built to produce a usable image in very few steps, and its Apache 2.0 license lets you ship the outputs in a commercial product without negotiating terms.
Expect clean composition and decent lighting. Expect weaker small text and more anatomy slips than the $25 and $40 tiers.
P-Image is Pruna's text-to-image model, and Pruna's page states the price plainly: $0.005 and about one second per image. That is $5 per 1,000. Z Image Turbo lands at the same $5 on fal.ai at one megapixel. The extra $2 over the $3 floor buys a different look, not a guaranteed upgrade, so run both on your own prompts before you pick.
Full disclosure: every photograph in this article was generated with P-Image. At Pruna's listed rate, the ten images cost five cents.
Step up one tier and prices jump by a multiple, not a few cents. FLUX Dev costs $25 per 1,000, about eight times the $3 floor. Nano Banana at $39.80, FLUX 1.1 Pro and Recraft V3 at $40 form a plateau near 13 times the floor. Nano Banana 2 at $80 and Ideogram v3 Quality at $90 sit above that, and Nano Banana Pro at $150 tops the list at fifty times the floor price.
Paying more is not wasteful by default. Premium tiers earn their price on text rendering, strict prompt following and consistent faces. The mistake is paying that rate for work a $3 model handles fine, like placeholders, thumbnails and background art.
Unit prices look tiny until you multiply them. This table runs each price point across three monthly volumes.
| Cost per 1,000 images | 1,000 per month | 10,000 per month | 100,000 per month |
|---|---|---|---|
| $3 | $3 | $30 | $300 |
| $5 | $5 | $50 | $500 |
| $25 | $25 | $250 | $2,500 |
| $40 | $40 | $400 | $4,000 |
| $90 | $90 | $900 | $9,000 |
| $150 | $150 | $1,500 | $15,000 |
At 10,000 images, every option except the most expensive stays under $1,000. A store refreshing its catalog pays $30 to $50 at the bottom tier and $900 to $1,500 at the top. The gap between the $3 tier and the $40 tier is $370 a month. That is real money for a hobby project and a rounding error for a funded product, so let quality drive the choice unless you are bootstrapping.
Here the unit price becomes strategy. 100,000 images at $3 per 1,000 is $300 a month. The same volume at $90 is $9,000, a difference of $8,700 a month or $104,400 a year.
At this scale a hybrid pipeline pays off. Send bulk, low-stakes images to the cheap tier and route only hero images and text-heavy graphics to a premium model. If 90% of the volume stays at $3 and 10% goes to a $90 tier, the blended price is 0.9 × 3 + 0.1 × 90 = $11.70 per 1,000, about $1,170 for 100,000 images.
A price list tells you what a generation costs. It does not tell you what a usable image costs, and that second number is the one on your invoice.
Use this formula: cost per 1,000 usable images = price per 1,000 divided by acceptance rate. Suppose a $3 model gives you an acceptable result 55% of the time. Your real cost is $3 ÷ 0.55 = $5.45 per 1,000 usable images. A $40 model with a 95% acceptance rate costs $40 ÷ 0.95 = $42.11. The cheap model still wins by almost eight times.
The $3 model only loses its advantage when its acceptance rate falls to about 7%, because $3 divided by $40 is 7.5%. Below that, you are paying for retries instead of images.
Better prompts raise the acceptance rate, and a better acceptance rate is a discount on every image after it. A fast language model such as Gemini 3.5 Flash can expand a short product description into a detailed, consistent prompt template. Test the effect on the same 20 prompts: if acceptance climbs from 55% to 70%, the $3 tier drops from $5.45 to $4.29 per 1,000 usable images.
You can also automate the rejection step. A vision-capable language model such as Qwen3.7-Plus can read each image and flag broken hands, garbled text or off-brief compositions before a human sees them. Price that reviewer call first, because a review that costs more than the image defeats the purpose.
Generating at four times the pixels multiplies a per-megapixel bill by about four. A cheaper route is to generate at one megapixel, review, and upscale only the accepted images. Real-ESRGAN is a 4x upscaler in the super resolution category. Check its per-run price before assuming it beats native high resolution, because some upscalers cost more than the base image they improve.
Throughput is a cost too. Use this formula: images per hour = concurrent jobs × 3,600 ÷ seconds per image. With 5 concurrent jobs and 10 seconds per image, you get 1,800 images an hour. With 1 second per image, the same 5 jobs produce 18,000. A cheap API with a slow queue can force you to buy a second provider just to hit a deadline, so test latency under load, not only on a single call.
Running an open model on your own GPU removes the per-image meter and adds a monthly rent plus maintenance. The break-even formula is simple: monthly GPU cost ÷ price per image = images per month at which renting wins. As a hypothetical, say a dedicated GPU costs $400 a month. At $0.003 per image, that is $400 ÷ 0.003 = about 133,000 images a month before the rented server pays off, and the figure ignores the hours you spend on drivers, scaling and uptime. Below that volume, a hosted API is cheaper and carries no pager duty. Replace the $400 with a real quote from your own cloud before you decide.
Thumbnails, social previews, placeholders and background textures belong in the $3 to $5 tier. Nobody inspects them at full size, and a retry costs a fraction of a cent. The decisions get harder in the two cases below.
A shop with 5,000 products and four angles per product needs 20,000 images. At $5 per 1,000 that is $100. At $40 per 1,000 it is $800. The right answer depends on how closely buyers scrutinize the picture. Lifestyle backgrounds tolerate the cheap tier. Hero shots that must show true leather grain or ceramic glaze are worth testing on a $25 or $40 model first.
Text is where cheap models struggle and premium models pay for themselves. Ideogram v3 Quality costs $0.09 per image on Replicate. If FLUX Schnell needs 30 attempts to spell a headline correctly, that is 30 × $0.003 = $0.09, the same as one premium attempt, plus your time. For 200 headline graphics, the premium route costs $18 in total. Spend there, and keep the volume work on the floor tier.
Developers who want a flat fee instead of a meter have another option. PicassoIA exposes a Replicate-style API at https://api.picassoia.com/v1 with four models: picassoia/picassoia-image for text to image, picassoia/picassoia-image-editor-pro for editing, picassoia/picassoia-video and picassoia/seedance-2.5-lite for video. The image models are listed as PicassoIA Image and PicassoIA Image Editor Pro.
Authentication uses a Bearer token that starts with pia_sk_. Jobs are asynchronous: you create a prediction, then poll it.
curl -s -X POST https://api.picassoia.com/v1/models/picassoia/picassoia-image/predictions \
-H "Authorization: Bearer $PICASSOIA_TOKEN" \
-H "Content-Type: application/json" \
-d '{"input": {"prompt": "a lighthouse at sunset, oil painting", "aspect_ratio": "16:9"}}'
The response carries a prediction ID. Send a GET request to /v1/predictions/{id} with the same header, waiting eta.next_poll_in_seconds between calls, until status reads succeeded, failed or canceled. Results arrive in the output field, and a failure explains itself in error.
The documented limits are 5 concurrent predictions per account, shared across tokens and MCP connections, a 10 MB request body, 4,000-character prompts and a 3-hour timeout. Five concurrent jobs is the real ceiling: use the throughput formula above and you can estimate your daily maximum.
On cost, the API page says predictions are currently free and use no credits, and that the API is available to accounts on the Infinite plan, which the pricing page lists at €24 per month on annual billing. The pricing page's feature list shows API access on more than one tier, and the wording can change, so confirm the current terms at picassoia.com/en/pricing before you build on it. As a break-even check, and ignoring the euro and dollar difference, a €24 flat fee beats a $3 per 1,000 meter once you pass about 8,000 images a month (24 ÷ 3 = 8).
Price lists are a starting point, not a verdict. The quickest way to find your cheapest usable image is a small experiment, and it takes under an hour.
Four habits keep the number low once you ship:
You can run that experiment today without writing a line of integration code. Open PicassoIA, try FLUX Schnell, P-Image and Ideogram v3 Quality side by side with your own prompts, and see which tier clears your quality bar. Then bring the winner into your API budget with real numbers instead of guesses.
