How much does an image generation API cost per image?
Per-image rates run from roughly a quarter of a cent on the fastest distilled models to about $0.25 on premium models. Most production text-to-image work sits between $0.02 and $0.05 an image. Models bill either per generated image or per megapixel of output, so a per-megapixel rate has to be multiplied by your resolution before you can compare it to a flat per-image rate.
What is the cheapest image generation API?
The cheapest rates are on distilled and turbo variants — sub-cent per image, or half a cent per megapixel. They are excellent for iteration, bulk variations and thumbnails, and weaker on prompt adherence and in-image text, so the practical answer is to iterate on a cheap model and spend on the final render. There is no subscription, so the cheap tier is genuinely cheap rather than a plan minimum.
Is image generation priced per image or per megapixel?
Both, depending on the model — and the two are not comparable without a resolution. A $0.02-per-megapixel model costs about $0.021 for a 1024×1024 image (1.05 MP) but about $0.084 at 2048×2048 (4.19 MP), where a flat per-image model would not move at all. Check the unit before comparing two rates.
Do I need a subscription to use an image generation API?
No. Billing is per request at each model’s published rate, snapshotted at request time, drawn from prepaid credits and exact to six decimals — no monthly plan, no seat, no per-request minimum, and no allowance that resets at the end of a billing month. You pay for the images you generate and nothing else.
Claude Code has no image generation on its own — connect the ModelRunner MCP and it calls image models as native tools. A real 29-image iOS batch shows how.
Image generation API pricing spans roughly two orders of magnitude. The same prompt costs about a quarter of a cent on one text-to-image model and twenty-five cents on another — close to a 100× spread — and the gap is not simply the difference between a bad model and a good one. Most published comparisons make it harder to read than it needs to be, because they flatten two incompatible units into a single column: some models bill a flat rate per generated image, others bill per megapixel of output, and a rate in one unit tells you nothing about a rate in the other until you fix a resolution. Below are the real per-model rates for 15 models, the conversion arithmetic, and how to choose.
What an AI-Generated Image Actually Costs in 2026
≈$0.0025
Lowest per-image rate in the table below
$0.25
Highest per-image rate among the 15
$0.005
Lowest per-megapixel rate
Those three figures are the shape of the whole market, and the third one is why the first two can mislead. A text-to-image model bills in one of two units. A per-image model charges a flat rate for each generated image, whatever size you asked for. A per-megapixel model charges for the pixels it actually produced, so the same model costs four times as much at 2048×2048 as it does at 1024×1024. Those two units are not interchangeable, and a table that prints them in one column — as most price roundups do — is comparing numbers that do not measure the same thing. The range above is raster output only; vector models emit SVG rather than pixels and are priced as a different product entirely.
Image Generation API Pricing Compared: 15 Text-to-Image Models
Every rate in this table is read live from the ModelRunner catalog when the page loads — the same records the billing system charges against. No price here is transcribed by hand, so the table cannot go stale between the day this was written and the day you read it. The first ten rows are priced per generated image, the last five per megapixel of output.
Four things worth reading out of that table:
The sub-cent tier is real. Distilled, few-step models generate a 1024px image for a fraction of a cent. That is not a promotional rate or a trial allowance — it is the standing price.
Most production work sits in a narrow mid-band. Between about two and five cents an image you get the current general-purpose flagships, and the spread inside that band is far smaller than the spread across the whole table.
Above the mid-band you are buying instruction-following, not resolution. A flat per-image rate does not move with output size, so the premium rows are not charging you for extra pixels — they are charging for stricter prompt adherence, reliable in-image text, and fewer discarded attempts.
Some rows read from $…, and that is not hedging. Those models price by output size or tier, so the rate shown is the floor and a larger render bills above it. Treat a from row as a range, never a flat rate.
These 15 are the ones people actually compare. The complete, always-current list — nearly 40 text-to-image models — is the full text-to-image price table.
Per Image vs Per Megapixel: The Comparison Everyone Gets Wrong
This is the arithmetic every price roundup skips, and it decides which model is cheaper more often than the headline rate does. A megapixel is one million pixels, so an output size converts to a bill by multiplication: width × height ÷ 1,000,000 × the rate. Run that for one per-megapixel rate against one flat per-image rate at three common output sizes and the ranking inverts inside the table.
Output size
Megapixels
At $0.02 per megapixel
At a flat $0.03 per image
1024×1024
1.05 MP
$0.021
$0.030
1024×1536
1.57 MP
$0.031
$0.030
2048×2048
4.19 MP
$0.084
$0.030
Two illustrative rates costed at three common output sizes. The per-megapixel side wins at 1024×1024, has already lost by 1024×1536, and costs nearly three times as much at 2048×2048 — while the flat rate never moves.
The crossover is not a property of either model — it is just the flat rate divided by the per-megapixel rate. For the pair above that lands at 1.5 megapixels: below it the metered side is cheaper, above it the flat side is. Which means “what is the cheapest image generation API” is not a question with an answer until you say how big the image is.
A worked example from this very page: the cover image at the top was generated on Z-Image Turbo, a per-megapixel model, at a 16:9 1K frame — 1.06 megapixels — and the request was charged $0.005284. That is the model’s per-megapixel rate multiplied by the pixels it produced, billed to six decimal places. Ask the same model for a 4K frame and that figure moves with the pixel count; ask a flat per-image model and it would not move at all.
The Cheapest Image Generation API Models (Under $0.02 an Image)
If your bottleneck is volume rather than polish, the bottom of the table is worth knowing by name. These three are the genuinely cheap end, and each is cheap for a different reason.
SDXL Lightning 4-step — ≈$0.00251692 per image. A distilled SDXL that reaches a 1024px image in four denoising steps. The cheapest way to see an idea rendered at all, and fast enough to sit behind an interactive feature.
Z-Image Turbo — $0.005 per megapixel. Turbo-class and billed per megapixel, so a small exploration costs proportionally less instead of a flat fee. At around 1 MP that still lands in sub-cent territory, and it drops further as you shrink the output.
Grok Imagine Image — $0.02 per image. Sits right at the top of the band, with aspect-ratio presets and up to four images per request — each one billed, so a four-image call is four charges.
Now the honest part, because a cheap-models section that only lists upsides is not useful. The cheap tier trades away the two things that most often force a re-run: strict prompt adherence on a long, multi-part description, and legible text rendered inside the image. Use it for iteration, thumbnails, bulk variations, moodboards and A/B fodder — the work where you need twenty results and will keep one. Do not use it for a hero asset, a page you will not revisit, or anything with words in it. The saving is real, and it is a saving on exploration, not on the final render.
When Paying More Per Image Is Cheaper Overall
The rate on the table is the price of one attempt. What you actually spend is the rate multiplied by the number of attempts it takes to get something you keep — and those two move in opposite directions. A model at four times the rate that lands the brief on the first try beats a cheap model you prompt five times and still nudge in an editor.
The arithmetic is worth doing explicitly, because intuition gets it wrong. Four attempts at three and a half cents costs more than one attempt at a rate three times higher, and that is before the minutes you spent judging the four. The premium rows earn their price precisely where re-runs are expensive: layouts with copy in them, brand-critical assets, anything with a specification longer than a sentence. The cheap rows earn theirs where re-runs are free because you wanted variety anyway.
There is a third lever that changes this calculus — explore at low resolution on a per-megapixel model, then upscale only the keeper — which we work through in our guide to 4K AI images without overpaying. It stacks with everything above rather than replacing it.
Pay-Per-Use vs Subscription: What “No Subscription” Actually Changes
Every rate on this page is metered per request. That is a genuinely different arrangement from the subscription and credit-pack model most consumer AI image tools sell, and the difference is not mainly about headline price — it is about what you can know and what you can waste.
The rate is published before you call. You can read a model’s price, do the arithmetic above, and decide — rather than discovering the cost after the fact, or translating it out of an opaque credit unit.
The price is fixed at request time. Each request carries its own pricing snapshot taken when it was created, so a rate change mid-run cannot reprice work already submitted.
Billing is exact, not rounded into buckets. Charges come off prepaid credits to six decimal places — the cover image on this page cost $0.005284, not “one credit”.
Nothing is reserved on your behalf. No seat, no monthly floor, no per-request minimum, and no allowance that resets at the end of a billing month whether you used it or not.
The practical consequence is that the cheap tier is actually cheap. Under a monthly plan, a sub-cent model saves you nothing, because the floor you pay is the plan — the marginal cost of an image is only real when there is no plan for it to disappear into. Current rates for every model live on the pay-per-use pricing page.
If you are weighing platforms rather than models, our platform comparisons cover that side of the decision.
How to Check a Model’s Price Before You Call It
A price table in a blog post is a snapshot, even a live one. Three ways to read a rate from the source, in increasing order of how little context-switching they cost:
1
Read it off the model’s own page
Every model in the catalog publishes its own rate, its pricing unit, and its input schema on its page. That page is the authority for a single model.
2
Scan every rate at once on the pricing page
Pricing lists all of them together, grouped by unit — the fastest way to answer “what else is in this band”.
3
Let your assistant read the price and run the model
Through the ModelRunner MCP server, an AI assistant can look up a model, read its live rate, and call it as a native tool. One config entry, and pricing stops being a tab you switch to.
Terminalbash
claude mcp add --transport http modelrunner https://mcp.modelrunner.run/mcp
That single line covers the third route for Claude Code; other MCP clients take the same URL. Whichever route you pick, the number you read is the number you are charged.
How the prices on this page are produced
The comparison table is not a hand-maintained list. Each row declares a catalog endpoint, and the page resolves that endpoint against the live catalog when it renders, formatting the rate with the canonical catalog price helper — the same one behind the catalog listings. The prices are therefore generated at read time rather than transcribed at writing time, which is the whole point of publishing them this way. A model’s own page remains the authority for its exact tier breakdown.
The 15 rows are curated rather than exhaustive: they span the full price range, cover both pricing units, and include one tiered row so the from $… shape is visible. The catalog holds 150+ public models in total, of which nearly 40 are text-to-image; the complete list is on the text-to-image hub.
The cover image was generated on ModelRunner for this post rather than licensed:
