Image generation is typically billed per image, out of a credit balance, or through a monthly subscription with a generation cap. Which one is cheaper depends almost entirely on how many attempts it takes you to get a picture you will use.
Budget from your reject rate, not from the count of finished images you need. That is the whole idea, and everything below is how to apply it.
How is image generation typically priced?
Per image. You pay a fixed amount for each generated image. Costs scale linearly and predictably. A batch of four costs four times a batch of one, whether you keep all four or none.
Credits. You buy a balance and each generation draws from it, often at different rates for different models or sizes. This is per-image pricing with a prepayment and an expiry date attached. The thing to check is whether unused credits expire, because that turns a balance into a subscription you did not sign up for.
Subscription. A monthly fee including some number of generations, sometimes with slower or unlimited "relaxed" generation past the cap. Cheap per image if you use the allowance, expensive if you do not.
Many services combine these: a free allowance, then per-image billing, or a subscription with credit top-ups. What you are looking for is what happens at the boundary, which is where the surprises live.
Why is your cost higher than you think?
Because you do not use most of what you generate. A usable image is the survivor of several attempts, and the discarded ones cost exactly the same as the one you kept.
The multiplier is the number that decides your bill. If a finished image takes you eight generations, your per-image price is effectively eight times the sticker price. Someone producing a specific composition to a brief will have a much higher multiplier than someone who needs any decent abstract background.
Nobody can tell you your multiplier in advance. What you can do is measure it: run a small batch of the work you intend to do, count total generations, count keepers, and divide. Do this before committing to a plan, not after.
Two things reliably lower the multiplier. Generating a batch from one prompt rather than one at a time, because you are sampling a range instead of rerolling. And iterating deliberately, changing one element at a time, rather than rewriting the prompt wholesale between attempts.
Which pricing model fits which work?
Match the pricing to the shape of your usage rather than to the headline rate.
Per image or credits suit irregular work, low volume, spiky projects, and anything where you might not touch it for a month. You pay for what you use and nothing when you do not. They also suit teams who need to attribute cost to a project, since each generation is a line item.
Subscriptions suit steady daily volume above the break-even point, and people who want a predictable bill more than a minimal one. If you generate every day, the per-image equivalent of a subscription is usually lower.
The break-even is straightforward: divide the monthly fee by the per-image price elsewhere, and that is how many images you have to generate for the subscription to win. Compare it against your measured volume, including rejects.
The trap is seasonal work. A subscription bought during a busy month keeps charging through the quiet one. If your work comes in bursts, per-image pricing usually wins even when it looks more expensive per unit.
What costs are not on the pricing page?
Several, and they are the ones that change the answer.
Retention. Many services delete generated images after a window. If you need a library, the cost of keeping it is yours: storage, and the discipline to download things. Discovering this after a purge is expensive in a way no rate card shows.
Upscaling. If you need large images, the enlarge step is often separately priced or a separate tool. Budget it as part of the finished-image cost.
Failed and refused generations. Ask whether a job that fails or gets refused still draws down your balance. Policies differ and it matters at volume.
Size and model tiers. A larger size or a better model frequently costs more per image. A quoted price is usually the cheapest combination.
The compute underneath the price. Image generation is at the expensive end of machine learning inference, and that is what the per-image rate is paying for. Power Hungry Processing: Watts Driving the Cost of AI Deployment? by Luccioni, Jernite, and Strubell, published at FAccT 2024, measured the energy and carbon needed for 1,000 inferences across model categories and found that "multi-purpose, generative architectures are orders of magnitude more expensive than task-specific systems for a variety of tasks, even when controlling for the number of model parameters." If a provider's price looks far below everyone else's, that gap is worth understanding before you build on it.
The people time. At small volumes this dominates everything. If a usable image takes forty minutes of prompting, the generation fee is not what the picture cost you.
How do you estimate a monthly bill?
Work it out from measured inputs rather than from intentions:
Count finished images you need per month. Be concrete, per project or per channel.
Measure your multiplier on a sample of the work you will do. Total generations divided by keepers.
Multiply. Finished images times multiplier is your true generation count.
Add the batch habit. If you generate four at a time, round up to whole batches.
Price it against each model at that volume, including the free allowance.
Add the extras above that apply, upscaling especially.
Then check the boundary. What happens when you exceed the plan? A hard stop, overage billing, or a downgrade to a slower queue are three very different outcomes.
Step seven is where people get caught. A plan that silently drops you to a worse model at the cap will not show up in your bill, it will show up in your output quality, and it takes a while to notice.
What should you check before committing?
Beyond price, these decide whether the service is usable for the work:
Is there a free allowance, and how does it reset? A rolling window and a calendar month behave differently. A rolling window means capacity returns gradually rather than all at once on the first.
Does an account or a card come first? Some services require payment details before you can evaluate output quality.
Can you generate batches? This changes your multiplier, which changes your bill more than the per-image rate does.
What are the commercial usage rights? Whether you can use the output commercially is a licensing question, separate from what you were allowed to generate. Read both.
Is the model named and stable? A service that silently swaps the model behind an endpoint will change your results, and your prompts, without notice.
What is logged and reviewed? Prompts are generally stored and subject to review. This is standard, and worth knowing rather than assuming otherwise.
The cheapest thing you can do
Get better at the loop before you optimize the rate. Halving your multiplier saves more than moving to a service that is twenty percent cheaper per image, and it is free.
That means generating in batches, changing one thing at a time, keeping the prompts that worked, and picking the right aspect ratio before iterating rather than after. The people who complain that image generation is expensive are usually generating one at a time and rewriting the prompt from scratch between attempts.
Where Carpathian fits
Carpathian AI prices image generation per call, after an allowance that every free account carries on a rolling window rather than a calendar reset, so capacity comes back gradually instead of all at once on the first of the month. The current per-image rate for each model is published on the model rate card, which reads from the same catalog the dashboard bills against, so the advertised figure and the charged figure are the same one. Chatting with the text models on the web is free and unmetered, which is a different thing from image generation and worth not confusing.
AI Image Gen Costs, and How to Budget for It | Carpathian
