Generate imagesLarge Language Models

OpenAI Image Generation API Pricing: Cost per Image, Free Options and Alternatives

What does one image cost through OpenAI's API? This article breaks down per image prices for GPT Image 1, 1.5 and 2, shows how quality and size move the bill, checks for free options, and compares cheaper alternatives you can run right now.

OpenAI Image Generation API Pricing: Cost per Image, Free Options and Alternatives
Cristian Da Conceicao

You open the billing dashboard after a weekend of testing an image feature, and the number is higher than your coffee budget for the month. It happens to almost everyone who wires up an image model for the first time. The reason is simple: OpenAI does not really charge "per image" the way a stock photo site does. It bills tokens, and the price you see per picture is a conversion from those tokens. Once you see how that conversion works, the whole pricing page stops feeling random.

This article lays out what each OpenAI image model costs per image, how quality and resolution change the total, whether anything is actually free, and which alternatives give you comparable results for less. Prices below come from published rates and third party trackers as I found them in November 2026. They shift, so treat every figure as an estimate and check OpenAI's own pricing page before you commit a budget.

Developer typing on a laptop at a wooden desk with coffee and notepad

What OpenAI Charges Per Image

Token billing in plain terms

OpenAI's GPT Image family bills output tokens. Every image you generate is turned into a number of tokens, and that number depends on the quality level and the canvas size. You pay a rate per million tokens, and the "cost per image" shown in tables is just that math done for you.

There is also an input side. Your text prompt costs text tokens, and if you send a reference photo for editing, that photo costs image input tokens. For a plain text to image call, the output tokens are the big line item and the prompt is almost a rounding error.

💡 Tip: When two pricing pages disagree by a few cents, check whether one of them includes input tokens and the other does not. That alone explains most of the mismatch.

GPT Image 1.5 price points

GPT Image 1.5 is the model most teams reach for when they want the sweet spot between cost and fidelity. OpenAI also says its image inputs and outputs run about 20 percent cheaper than the original GPT Image 1. For a standard square 1024 by 1024 image, the figures I found were roughly:

QualityApprox. cost per 1024 x 1024 image
Low$0.009
Medium$0.034
High$0.133

Larger canvases cost more. A portrait or landscape frame sits a bit above the square price, and the exact rate per quality and size is listed on the official model page.

GPT Image 1 and the mini tier

The original GPT Image 1 is still available and still priced higher. At the same square size, trackers list about $0.011 for low, $0.042 for medium and $0.167 for high. If you are on it today, moving to 1.5 is the easiest saving on this page.

There is also a mini version aimed at cheap, high volume work. I could not confirm a reliable per image number for it, so I will not guess one. Check the current rate before you build a forecast around it. You can test how it behaves on PicassoIA with GPT Image 1 Mini and compare the output with GPT Image 1 and GPT Image 1.5.

GPT Image 2 token rates

The newest generation, GPT Image 2, is where numbers get messy. Third party sources report token rates of about $5 per million for text input, $8 per million for image input and $30 per million for image output. Converted to a square image, that lands anywhere from around $0.006 at low quality up to about $0.21 at high quality.

Resellers advertise their own flat prices on top of that, some well under the official rate and some far above it. Those are not OpenAI prices. If you buy through a reseller, you are paying for convenience, uptime or a bundle, so compare it against the official token math rather than against another reseller.

How Quality and Size Change the Bill

Macro of hands sliding silver coins beside a receipt and calculator

Low, medium and high tiers

Quality is the biggest lever you control. On GPT Image 1.5, high quality costs roughly fifteen times more than low for the same square canvas. Medium sits in the middle at around four times the low price.

Here is what that gap buys you in practice:

  • Low: fine for thumbnails, moodboards, placeholder art and quick prompt testing. Fine detail like hands, small text and fabric weave can look soft.
  • Medium: the everyday setting for blog headers, social posts and product mockups.
  • High: worth paying for when the image is the product, such as a hero banner, an ad creative or a print.

Three photographic prints of a mountain lake in rising quality on a table

Resolution and aspect ratio

Token count grows with pixel area. A 1024 by 1024 image is the baseline, and wider or taller formats cost a bit more. Going to a large 2K or 4K output, where the model supports it, can multiply the cost several times over.

A practical habit: generate at the smallest size that still works for the final use. If the image will sit in a 600 pixel wide blog column, paying for a huge canvas is wasted money.

Edits and image inputs

Editing calls cost more than plain generation because you are also paying for the reference image as input tokens. If your workflow sends three reference photos with every request, that input cost can rival the output cost at low quality. Send only the references the model actually needs.

Monthly Cost at Real Volumes

What 1,000 and 10,000 images cost

Using the GPT Image 1.5 and GPT Image 1 square image estimates from above, here is how the numbers stack up as you scale:

VolumeGPT Image 1.5 lowGPT Image 1.5 mediumGPT Image 1.5 highGPT Image 1 high
1,000 imagesabout $9about $34about $133about $167
10,000 imagesabout $90about $340about $1,330about $1,670
100,000 imagesabout $900about $3,400about $13,300about $16,700

The takeaway is not the exact dollar value. It is that quality choice matters more than model choice at almost every volume. Moving a workload from high to medium saves far more than switching vendors.

A small shop example

Shop owner packing a ceramic mug into a cardboard box at a workbench

Picture a small store with 400 products that wants three lifestyle images per product. That is 1,200 images. At medium quality on GPT Image 1.5 you are looking at roughly $40 to $45 for the whole catalog. At high quality it jumps to around $160. Add a second pass because the first generation of a few images missed the brief, and you can easily double the figure.

This is why most production teams settle on a draft and final rhythm. Draft at low, pick winners, regenerate only the winners at high.

💡 Tip: Budget for retries. In real projects, expect to generate 1.5 to 3 images for every one you publish.

Free Options Worth Trying

Student working on a laptop at a long table in a quiet library

Does the API have a free tier?

From what I found, OpenAI's image API does not come with a free image generation allowance. You add credit to your account, and each call draws from that balance. Some accounts have received promotional credits over time, but those are not a standing free tier, so do not plan a product around them.

Free generation inside ChatGPT

OpenAI has offered limited image generation inside ChatGPT for free accounts, with daily caps that change from time to time. That is a handy way to test prompts by hand. It is not an API, you cannot automate it, and you should not assume the limits will stay the same next month.

Free and cheap models on PicassoIA

If your goal is testing ideas without watching a meter, PicassoIA gives you a long list of text to image models in one place, including several from OpenAI. You can try GPT Image 2.5 Flare for fast results and GPT Image 2.5 Sunburst for generation and editing. DALL-E 3 is also there if you want to compare the older model against the newer ones. Use the model pages to see current availability and limits, since those change.

Cheaper Alternatives Today

Designer comparing several printed photographs against daylight in a studio

Models worth comparing

OpenAI is not the only way to get a photorealistic image. These models are all on PicassoIA and are worth a side by side test with your own prompts:

ModelBest forLink
Seedream 4.5Sharp, high resolution images from textSeedream 4.5
Nano Banana ProDetailed 4K output and image editingNano Banana Pro
Nano Banana 2 LiteQuick, low cost draftsNano Banana 2 Lite
Imagen 4Photorealistic scenes and portraitsImagen 4
Imagen 4 FastSpeed on large batchesImagen 4 Fast
Flux SchnellVery fast generation for iterationFlux Schnell

When OpenAI still wins

Price is not the only variable. GPT Image models tend to follow long, detailed prompts closely and render text inside images better than many rivals. If your work involves posters, packaging mockups or infographics with real words, paying more can save you retries, and retries are a cost too.

For pure photography style output, product shots or lifestyle scenes, the alternatives above are often close enough that the cheaper option wins on total spend.

How to Use GPT Image 1.5 on PicassoIA

Hands resting on a laptop keyboard in a bright workspace

If you want to see what a GPT Image result looks like before you commit API credit, run a few prompts through GPT Image 1.5 on PicassoIA. It is a quick way to judge quality with no integration work.

Step by step

  1. Open the GPT Image 1.5 page and sign in.
  2. Write a prompt with a clear subject, setting, lighting and camera feel. Something like "a ceramic mug on a linen tablecloth, soft window light from the left, 85mm lens, shallow depth of field" works well.
  3. Choose the options the page offers, such as aspect ratio, and a transparent background when you need a cutout.
  4. Generate, review, and refine the prompt rather than rerunning it unchanged.
  5. Download the result you like and compare it with the same prompt on another model.

Parameter tips

  • Be specific about lighting. "Soft north window light" gives more realistic results than "nice lighting."
  • Name the lens and the film look when you want photographic output.
  • Keep one idea per prompt. Long prompts with five competing ideas usually lose detail on all five.
  • Run the same prompt on two models. It takes two minutes and tells you whether a premium model earns its price for your content.

Ways to Cut Your Image Spend

Aerial view of a creative desk with a budget sheet, calculator and prints

Draft low, finish high

Run exploratory prompts at low quality. Only the winners get a medium or high rerun. This single habit often cuts a bill by more than half.

Cache and reuse

If a user requests the same image twice, serve the first one. Store the prompt, the settings and the output, and check for a match before you call the API again.

Let a language model write the prompts

Weak prompts waste money because they produce images you throw away. A language model such as GPT 5.6 Luna can turn a short idea into a detailed, photographic prompt in seconds. For harder briefs, GPT 5.6 Terra handles longer instructions. Better prompts mean fewer retries, and fewer retries are the cheapest saving of all.

Set hard limits

Put a monthly spend cap on your OpenAI account and add per user rate limits in your own app. A single bug in a loop can burn through a budget overnight.

Try It Yourself on PicassoIA

Photographer hanging a large framed landscape print on a gallery wall

Pricing tables only tell you so much. The fastest way to decide whether an image model is worth its cost is to run your own prompts and look at the results side by side. Open Picasso IA, pick two or three models from this article, such as GPT Image 1.5, Seedream 4.5 and Imagen 4, and feed them the same prompt. Keep the one whose output you would actually publish, then do the cost math for that model alone.

Ready to see more options? Browse every model on picassoia.com/en/all-models and start generating your own images today.

Share this article

Choose your language

FreeCreate AI Images and VideosFree, No Credit CardFree Generations to StartSeedance 2.5 Lite IncludedStart Free Now