The lowest and highest per-image list prices on the same API catalog belong to the same provider. On the Modellix image catalog measured September 15, 2026, one OpenAI route publishes a low end of $0.0054 per image and another publishes a high end of $0.2250 per image — a 41.7× spread inside a single vendor’s own lineup, on a single key. That is not a pricing quirk. It is the reason “AI image model API pricing comparison” is a harder question than it looks: a per-image price is usually a band, and a band is a parameter, not a number.
Here is what the whole catalog actually looks like as of September 15, 2026: 49 live image routes across 7 first-party providers — 22 text-to-image and 27 image-to-image — every one of them quoted per image, and 14 of the 49 publishing a range instead of a single figure. We measured all 49 from the individual model pages on that date, spot-checked 12 of them line by line against the live pages, and each table below names the model pages its figures were read from. Modellix is the aggregator whose catalog this is, so we have a commercial interest in this comparison; where the numbers favour us we have tried to say so plainly, and where they don’t, we have said that too — see the section on coverage near the end, where a competitor’s list is genuinely longer than ours.
What “per image” hides: units you cannot subtract from each other
An image-model API bill is assembled from four units, and they are not interchangeable:
- Per image — one flat figure per generated image, whatever the shape. This is how all 49 routes on this catalog are published, and it is also how most hosted open-weight models are quoted.
- Per megapixel — the bill scales with the pixel count of the output, so a 1024×1024 image and a 2048×2048 image are different products at different rates.
- Per second of GPU compute — common on platforms that rent hardware by the hour and pass through the runtime of the job. The same model can bill differently on two machines.
- Per output token — how OpenAI’s image models are metered upstream, per its own API pricing page. Token count depends on output size and quality tier, which is exactly why OpenAI’s own image routes appear in our table as a range rather than a point.
Two prices are only subtractable if they share a unit and the underlying call is the same shape. A per-megapixel figure and a per-image figure can both be correct, both be current, and still be incomparable: they are answering “how much per pixel” and “how much per picture”. This is the single most common way a pricing comparison misleads — not a wrong number, but a right number placed next to a non-comparable one.
The token-priced world has the same problem in a different costume: input and output tokens bill at different rates, cached reads bill at a third rate, and some models step up a price tier when the request crosses a context threshold. We cover that market separately in our LLM API pricing comparison, which prices 28 text models per million tokens. Nothing in that article applies to the table below, and nothing below applies there — which is precisely why the image catalog can be collapsed into one legitimate table while a mixed text-and-image table cannot.
The catalog these 49 rows were measured from, captured September 15, 2026. Each route’s price is visible on its own model page; the per-image unit is uniform across the group, which is what makes a single comparison column honest here.
The full price table: 49 routes, 7 providers, measured September 15, 2026
Everything below is a list price published per image, read from the individual model pages on September 15, 2026. A tilde means the route publishes a range. Nothing is averaged, and nothing is normalised to a reference resolution — where a route sells several configurations, we show the band the route itself publishes.
Text-to-image routes (22)
Route (provider/model) |
Billing unit | List price as of Sep 15, 2026 |
|---|---|---|
openai/gpt-image-2 |
per image | $0.0054~$0.1899/img |
openai/gpt-image-1.5 |
per image | $0.0117~$0.1800/img |
alibaba/z-image-turbo |
per image | $0.0135~$0.0270/img |
microsoft/mai-image-2.6-flash |
per image | $0.0195/img |
microsoft/mai-image-2.5-flash |
per image | $0.0200/img |
kling/kling-v3-t2i |
per image | $0.0224/img |
xai/grok-imagine-image |
per image | $0.0240/img |
alibaba/qwen-image-3.0 |
per image | $0.0270/img |
alibaba/wan2.7-image |
per image | $0.0270/img |
google/nano-banana-2-lite |
per image | $0.0306/img |
bytedance/seedream-5.0-lite |
per image | $0.0350/img |
google/nano-banana |
per image | $0.0351/img |
microsoft/mai-image-2.6 |
per image | $0.0389/img |
google/nano-banana-2 |
per image | $0.0405~$0.1359/img |
xai/grok-imagine-image-2.0 |
per image | $0.0480~$0.0960/img |
microsoft/mai-image-2.5 |
per image | $0.0481/img |
xai/grok-imagine-image-quality |
per image | $0.0600~$0.0840/img |
alibaba/qwen-image-3.0-pro |
per image | $0.0675/img |
alibaba/wan2.7-image-pro |
per image | $0.0675/img |
bytedance/seedream-5.0-pro |
per image | $0.0810/img |
microsoft/mai-image-2.5-pro |
per image | $0.1085/img |
google/nano-banana-pro |
per image | $0.1206~$0.2160/img |
Every figure above is the list price shown on that route’s own model page on www.modellix.ai, read on September 15, 2026 — for example openai/gpt-image-2, google/nano-banana-pro, bytedance/seedream-5.0-pro and microsoft/mai-image-2.6. img counts one output image. A tilde between two figures is the route’s own published range, reproduced without averaging.
Image-to-image and edit routes (27)
Route (provider/model) |
Billing unit | List price as of Sep 15, 2026 |
|---|---|---|
openai/gpt-image-1.5-edit |
per image | $0.0135~$0.1800/img |
microsoft/mai-image-2.5-flash-edit |
per image | $0.0218/img |
microsoft/mai-image-2.6-flash-edit |
per image | $0.0220/img |
kling/kling-image-expansion |
per image | $0.0224/img |
kling/kling-image-o1 |
per image | $0.0224/img |
kling/kling-v3-i2i |
per image | $0.0224/img |
kling/kling-v3-omni-image |
per image | $0.0224~$0.0448/img |
alibaba/wan2.7-image-edit |
per image | $0.0270/img |
alibaba/qwen-image-3.0-edit |
per image | $0.0297/img |
google/nano-banana-2-lite-edit |
per image | $0.0306/img |
bytedance/seedream-5.0-lite-edit |
per image | $0.0350/img |
google/nano-banana-edit |
per image | $0.0360/img |
openai/gpt-image-2-edit |
per image | $0.0360~$0.2250/img |
xai/grok-imagine-image-edit |
per image | $0.0360/img |
google/nano-banana-2-edit |
per image | $0.0405~$0.1359/img |
microsoft/mai-image-2.6-edit |
per image | $0.0471/img |
microsoft/mai-image-2.5-edit |
per image | $0.0563/img |
alibaba/wan2.7-image-pro-edit |
per image | $0.0675/img |
kling/kolors-virtual-try-on-v1 |
per image | $0.0700/img |
kling/kolors-virtual-try-on-v1-5 |
per image | $0.0700/img |
alibaba/qwen-image-3.0-pro-edit |
per image | $0.0702/img |
xai/grok-imagine-image-quality-edit |
per image | $0.0720~$0.0960/img |
bytedance/seedream-5.0-pro-edit |
per image | $0.0810/img |
bytedance/seedream-5.0-pro-multi-reference |
per image | $0.0810/img |
xai/grok-imagine-image-2.0-edit |
per image | $0.0840~$0.1320/img |
microsoft/mai-image-2.5-pro-edit |
per image | $0.1167/img |
google/nano-banana-pro-edit |
per image | $0.1215~$0.2169/img |
Source: each route’s own model page on www.modellix.ai, read September 15, 2026 — the same pages the text-to-image table above cites, plus google/nano-banana-pro-edit and openai/gpt-image-2-edit for the two widest edit bands.
Provider summary
| Provider | Routes in this table | Lowest published per-image price | Highest published per-image price |
|---|---|---|---|
| openai | 4 | $0.0054 | $0.2250 |
| alibaba | 9 | $0.0135 | $0.0702 |
| microsoft | 10 | $0.0195 | $0.1167 |
| kling | 7 | $0.0224 | $0.0700 |
| xai | 6 | $0.0240 | $0.1320 |
| 8 | $0.0306 | $0.2169 | |
| bytedance | 5 | $0.0350 | $0.0810 |
Two things to notice before reading any of it as a ranking. First, the columns are per-image list prices, not invoices — the low end of a band is the cheapest configuration the route sells, not the configuration you are going to call. Second, the spread inside a single provider is often larger than the spread between providers: microsoft spans $0.0195 to $0.1167, google $0.0306 to $0.2169, and openai $0.0054 to $0.2250. “Which provider is cheaper” is the wrong question to ask of this table, because the provider is not what sets the number — the route and its parameters are.
Why 14 of the 49 prices are a range, not a number
Slightly more than a quarter of this catalog — 14 of 49 routes, or 28.6% — publishes a range. That is the single most useful fact in the article, because it is the fact a single-figure comparison table cannot express.
The 14 band routes
| Route | Published list price | Band width |
|---|---|---|
openai/gpt-image-2-edit |
$0.0360~$0.2250/img | 6.3× |
openai/gpt-image-2 |
$0.0054~$0.1899/img | 35.2× |
openai/gpt-image-1.5 |
$0.0117~$0.1800/img | 15.4× |
openai/gpt-image-1.5-edit |
$0.0135~$0.1800/img | 13.3× |
google/nano-banana-pro-edit |
$0.1215~$0.2169/img | 1.8× |
google/nano-banana-pro |
$0.1206~$0.2160/img | 1.8× |
google/nano-banana-2 |
$0.0405~$0.1359/img | 3.4× |
google/nano-banana-2-edit |
$0.0405~$0.1359/img | 3.4× |
xai/grok-imagine-image-2.0 |
$0.0480~$0.0960/img | 2.0× |
xai/grok-imagine-image-2.0-edit |
$0.0840~$0.1320/img | 1.6× |
xai/grok-imagine-image-quality |
$0.0600~$0.0840/img | 1.4× |
xai/grok-imagine-image-quality-edit |
$0.0720~$0.0960/img | 1.3× |
kling/kling-v3-omni-image |
$0.0224~$0.0448/img | 2.0× |
alibaba/z-image-turbo |
$0.0135~$0.0270/img | 2.0× |
Band widths are arithmetic on the published figures above, not a vendor statement: the width column simply divides the route’s own high end by its own low end.
What moves a band is not volume. It is the parameters of the individual call — the output size and aspect ratio you ask for, the quality or tier setting, and for edit routes how many reference images you send. That is why the band is wide on some routes and nearly flat on others: google/nano-banana-pro spans 1.8×, while openai/gpt-image-2 spans 35.2× because the same route sells both a small low-tier output and a large high-tier one.
Where a vendor documents the parameter mapping on a page we can reach, we link to it — Google’s Gemini API pricing page lists the Nano Banana image tiers behind its ranges, and our own guide to Nano Banana Pro’s resolution tiers walks the same bands from the calling side. The principle is not Google-specific: ByteDance’s own ModelArk pricing page states it in one line — “A video generated at a higher resolution costs more than one generated at a lower resolution” — which is the same output-size multiplier that turns an image route’s rate card into a range.
The band is published verbatim, not derived: google/nano-banana-pro at $0.1206~$0.2160 per image on September 15, 2026 — and the same page exposes the parameters that select a point inside it, which is the standard the whole catalog table is held to.
A limit we are declaring rather than papering over: we did not individually confirm which parameter controls each band on all 14 routes. For some, the first-party documentation is explicit; for others we can see that a range exists and that it is selected per call, and we stop there. If you need the exact multiplier for one route before you commit a budget, pull its schema (there is a no-key way to do that, further down) rather than trusting a general statement about “resolution”.
Generate versus edit: six of twenty families charge nothing for the switch
Because generate and edit routes sit in the same table under the same key, a comparison nobody else can print becomes possible: what does it cost to move from making an image to changing one?
There are 20 model families in this catalog with both routes on sale. In 6 of them the edit route is priced identically to the generate route. In the other 14 there is a premium, from +$0.0009 to +$0.0360 on the published low ends.
| Model family | Generate route | Edit route | Edit premium (low end) |
|---|---|---|---|
openai · gpt-image-2 |
$0.0054~$0.1899/img | $0.0360~$0.2250/img | +$0.0306 (567%) |
openai · gpt-image-1.5 |
$0.0117~$0.1800/img | $0.0135~$0.1800/img | +$0.0018 (15%) |
google · nano-banana |
$0.0351/img | $0.0360/img | +$0.0009 (3%) |
google · nano-banana-2 |
$0.0405~$0.1359/img | $0.0405~$0.1359/img | $0 — same price |
google · nano-banana-2-lite |
$0.0306/img | $0.0306/img | $0 — same price |
google · nano-banana-pro |
$0.1206~$0.2160/img | $0.1215~$0.2169/img | +$0.0009 (1%) |
bytedance · seedream-5.0-lite |
$0.0350/img | $0.0350/img | $0 — same price |
bytedance · seedream-5.0-pro |
$0.0810/img | $0.0810/img | $0 — same price |
microsoft · mai-image-2.5 |
$0.0481/img | $0.0563/img | +$0.0082 (17%) |
microsoft · mai-image-2.5-flash |
$0.0200/img | $0.0218/img | +$0.0018 (9%) |
microsoft · mai-image-2.5-pro |
$0.1085/img | $0.1167/img | +$0.0082 (8%) |
microsoft · mai-image-2.6 |
$0.0389/img | $0.0471/img | +$0.0082 (21%) |
microsoft · mai-image-2.6-flash |
$0.0195/img | $0.0220/img | +$0.0025 (13%) |
xai · grok-imagine-image |
$0.0240/img | $0.0360/img | +$0.0120 (50%) |
xai · grok-imagine-image-2.0 |
$0.0480~$0.0960/img | $0.0840~$0.1320/img | +$0.0360 (75%) |
xai · grok-imagine-image-quality |
$0.0600~$0.0840/img | $0.0720~$0.0960/img | +$0.0120 (20%) |
alibaba · qwen-image-3.0 |
$0.0270/img | $0.0297/img | +$0.0027 (10%) |
alibaba · qwen-image-3.0-pro |
$0.0675/img | $0.0702/img | +$0.0027 (4%) |
alibaba · wan2.7-image |
$0.0270/img | $0.0270/img | $0 — same price |
alibaba · wan2.7-image-pro |
$0.0675/img | $0.0675/img | $0 — same price |
Read that table with two cautions, both of which cut against the obvious conclusion. The first: the low ends are not the same picture. A generate route’s $0.048 and its edit route’s $0.084 are different configurations — different output sizes, possibly different tiers — so “the edit route costs 75% more” is a statement about two price tags, not about two identical images. The second: a zero premium is a pricing decision, not a promise that editing is free in practice — an edit call that sends three reference images is a different call from one that sends a single prompt, and the input side is where that shows up. OpenAI’s image generation guide documents image-input tokens as a separate billed category, and our OpenAI image pricing breakdown works through what that does to a real per-image number.
One practical consequence for anyone building an edit-first product: the reference images you send have to get into the request somehow. On this platform that is the File API — an authenticated upload for image, video and audio assets that returns a URL you feed straight into a prediction call, at no charge for the upload itself, with assets retained for about 7 days. Build the pipeline assuming the reference expires, because it does.
What 10,000 images a month actually costs
Monthly cost is price per image times volume, so the only question that matters is which price. Here is the same 10,000-image month priced at four points on this catalog:
| If your route bills at | 1,000 images | 10,000 images | 100,000 images |
|---|---|---|---|
| $0.0054 (lowest published point) | $5.40 | $54 | $540 |
| $0.0360 (median of the published low ends) | $36 | $360 | $3,600 |
| $0.0810 (a flat per-image route, e.g. Seedream 5.0 Pro) | $81 | $810 | $8,100 |
| $0.2250 (highest published point) | $225 | $2,250 | $22,500 |
Now the part that costs people money. If a route publishes a band and you budget the midpoint, you are budgeting a shape nobody generates:
google/nano-banana-propublishes $0.1206~$0.2160. The midpoint is $0.1683, which over a 10,000-image month is $1,683. If you actually run the low configuration you spend $1,206; the high configuration is $2,160. The midpoint is off by $477 in both directions — roughly 40% of the low-configuration bill.openai/gpt-image-2publishes $0.0054~$0.1899. The midpoint, $0.09765, is 18× the low end. Budget $977 a month for 10,000 images and you may spend $54, or you may spend $1,899.
That is the whole argument for treating a band as a band. A midpoint is an average of two shapes, and nobody generates the average. Multiply a band by your own monthly volume and you have the only cost calculator this question needs — the arithmetic is never where the error is; which of the two ends you multiplied is. The same per-image logic is what drives ByteDance’s flat image rates to look deceptively simple by comparison: a flat rate is not a better deal, it is an easier number.
The cheapest published row is not the cheapest route for your job
Sorting this catalog by the low end of the range gives an order that is technically correct and practically useless. The lowest published points belong to routes sold as cheap, small, fast configurations; the highest belong to routes sold on output fidelity, typography, multi-reference editing and larger native resolution. They are not competing for the same job.
The most complete public image-price list we found on this query — Price Per Token’s image pricing comparison, which declares 145 image models across 5 providers and a lowest price of $0.0002 — is honest about its own rule in one line: “All prices normalized to 1024×1024 image generation” and “cheapest provider per model”. Normalising to one shape is what makes 145 heterogeneous models sortable at all. It is also what hides the thing this article is about: once every row is a point at one shape, a 3.4× internal band becomes invisible, and the top of the sorted list is a set of turbo models that were never going to produce the image you had in mind.
We are not going to tell you this catalog is the cheapest place to buy image generation, because it isn’t a claim we can support and the lowest-price hosts in this market are a different set of companies. What we can tell you is what each route costs, when we measured it, and which parameter moves it.
Illustrative view of the measured per-image price bands by provider — a schematic render, with bar endpoints deviating roughly 1–3% from the axis scale, so the tables above are the exact figures; bands measured September 15, 2026. The OpenAI span reaches both extremes of the whole catalog — which is the article’s opening finding, and the reason a single “cheapest provider” column misleads.
How to check what one call actually billed
Every number above is a list price. Your invoice is a different document, and the gap between the two is where a pricing article either helps or doesn’t. Two mechanisms close it.
Per-request cost in the request log. Modellix’s request log records the cost of each individual call alongside the model it resolved to, so you can take any single image you generated, look up what it billed, and compare that against the band you budgeted from. This turns the band from a guess into a measurement: run your intended configuration for a day, read the log, and you have your own per-image number instead of an estimate derived from someone else’s midpoint. The log also records the input parameters of each request (added June 17, 2026), which is what lets you correlate a price with the configuration that produced it.
Account-level discounts are not model-level prices. Modellix runs a first top-up discount of 10% and unlocks concurrency by spend tier (introduced June 23, 2026). That is a discount on a deposit, applied to your account balance — it is not a reduction of any route’s published per-image price, and any comparison that mixes the two is comparing a payment instrument to a rate card. The per-image figures in this article are the rate card.
How to re-pull these prices without an API key
A dated price table starts decaying the moment it is published, so the useful part of it is the way back to the source. Two endpoints do that here, and neither requires a paid account.
The model discovery endpoint returns the catalog — each route’s slug, type and documentation URL — so you can enumerate what is currently on sale rather than trusting a snapshot. Model descriptions are returned alongside, which is how you tell two similarly named tier variants apart. And the CLI’s schema command returns a model’s public request and response schema without an API key, which is the missing piece for the band question: it gives you the actual parameter names and permitted values for a route, so you can see for yourself what a given configuration will select rather than inferring it from a price range. The project’s own skill and plugin tooling moved to reading request bodies from that schema endpoint instead of guessing from documentation, which is the right default for anyone doing this work in an agent loop.
Image Model API Reference
Pull any image model's request schema and per-image pricing, then run it on your own key.
View DocsScope, limits, and where a competitor’s list is longer than ours
What this article covers: the 49 live image routes on this catalog as of September 15, 2026 — 22 text-to-image and 27 image-to-image — across 7 first-party provider catalogs (microsoft, alibaba, google, kling, xai, bytedance, openai). For scale within the same platform, the full model-page count on the same date is 178: these 49, plus 112 video routes (text-to-video 26, image-to-video 58, video-to-video 28) and 17 speech routes. Different definitions of “model count” produce different numbers, so we are naming the one we are using: live model pages with a published price, counted on September 15, 2026.
What it does not cover, stated plainly:
- Open-weight routes hosted by third parties. Stability’s and Black Forest Labs’ open-weight image models are not on this catalog, so a reader shopping specifically for a Flux or SDXL route should use a host that carries them — Price Per Token’s list is the more complete one for that job at 145 models, and it carries a visible last-updated stamp (September 14, 2026 when we checked). Our list is shorter. What it has instead is the first-party frontier routes (gpt-image-2, nano-banana-pro, seedream-5.0-pro, grok-imagine-image-2.0, mai-image-2.6, qwen-image-3.0) priced next to each other in one column, which is the set most 2026 image-API buyers are actually choosing between.
- Per-second and per-megapixel billing. It appears in this table only as an explanation of why cross-unit comparison fails. A host that charges for GPU seconds will not reproduce these numbers.
- Anything about throughput, rate limits or uptime. This is a pricing article; we are not attaching performance promises to it.
- Invoices. List prices only. See the verification section above for how to get from here to your own number.
Run the Image Catalog on One Key
Log in to compare these 49 image routes on your own prompts and read the per-call cost in your request log.
LoginHow to choose: match the band to the job
| If your main reason for choosing is… | Start with… |
|---|---|
| Lowest cost per image at scale, simple pipeline | A flat per-image route in the $0.0195–$0.0350 range (microsoft flash tiers, google’s lite tiers, bytedance Seedream Lite) |
| Prompt-accurate typography and instruction following | A frontier text-to-image route at $0.0810–$0.1359, then verify at your target resolution before scaling |
| Editing an existing image, keeping composition | An edit route from the six families where generate and edit are priced identically |
| Multi-reference composition from supplied assets | A multi-reference route, with the File API upload path built in from day one and its ~7-day retention assumed |
| Highest output fidelity, cost secondary | The upper band of nano-banana-pro or gpt-image-2-edit |
| You are unsure of your actual per-image cost | Any route, for one day, with the request log open |
Frequently Asked Questions About AI Image Model API Pricing
What is the cheapest AI image model API?
By published low end on this catalog, the lowest per-image figure is $0.0054 — but that is the cheapest configuration of the cheapest route, not the cheapest way to produce a given image. Sorting a price list by its low end ranks pricing configurations, not jobs, and the top of that list is dominated by small fast models that cannot do multi-reference editing or legible text. Measure your own configuration before treating any low end as your price.
How many image models does Modellix have?
49 live image routes as of September 15, 2026 — 22 text-to-image and 27 image-to-image — across 7 provider catalogs. That count is of model pages with a published price on that date; the full catalog including video and speech is 178. We state the measurement date because this number changes.
Are image API prices per image or per token?
On this catalog, per image — all 49 routes publish a per-image figure, 35 of them as a single number and 14 as a range. Upstream, OpenAI converts a token bill into a per-image range, and other hosts bill per megapixel or per GPU-second. Prices from different units cannot be subtracted from each other until you fix the unit.
Why do some models show a price range instead of one price?
Because the route sells more than one configuration. The range spans the cheapest and most expensive configurations the route offers — driven by output size, quality tier, and for edit routes the number of reference images. A range is not a discount schedule; it is a map of what your parameters can select.
Does Modellix offer free credits or a free tier?
No. The $1 registration credit was retired on August 19, 2026. New registrations no longer receive a complimentary credit; if you need trial quota, request it from support directly. Any page still promising free signup credit for this platform is out of date.
Is a per-image price the same as my invoice?
No. The published per-image figure is a list price; what you pay reflects the configuration you actually called, plus any account-level discount on deposits. The request log records the cost of each individual call, which is how you replace the estimate with your own number.
How often do these prices change?
Often enough that a table without a date is not worth much. Every figure here is as of September 15, 2026, and the catalog itself moves — the count, the tier lineups, and individual rates. Re-pull before you commit a budget: the model discovery endpoint lists what is on sale now, and the CLI’s schema command returns any route’s parameters without an API key.
Provider names, model routes and per-image prices reflect the Modellix catalog as of September 15, 2026 and change frequently — validate against the live model pages before committing a budget. Access image and video models, including the leading Chinese models, through a single API key at modellix.ai.
