There is no single number for what an AI image costs here, and the reason is structural rather than evasive: this platform runs two kinds of image model with two different bills behind them. Some run on GPUs in this building, where the marginal cost of one picture is electricity. The rest are HTTPS calls to Google, ByteDance, Alibaba and OpenAI, where the marginal cost is an invoice. Those two facts produce two wallets and two price lists, and this page is both of them in full.
An account carries two balances. Green credits (⚡) refill continuously toward a daily cap and pay only for models running on our own hardware. Gold credits (✦) are bought with money or granted monthly by a plan, and pay for everything a third party bills us for. They do not convert in either direction, and signing up grants zero gold.
That split is not a packaging decision, it is the shape of the underlying cost. A picture on our own GPU costs us a few seconds of electricity on a machine that is already paid for, so it can be generous and refill. A picture on a hosted endpoint costs us real money the instant the call is made, so it cannot. Every price below follows from which of those two things is happening.
What it costs: Own-GPU prices are quoted as "from". They are set at runtime, can carry a demand multiplier when the render queue is deep, and where a model has a hosted twin the run can be routed to it instead when the local lane cannot serve — Z-Image Turbo bills ✦2.5 in gold on that lane rather than ⚡1 in green. Treat ⚡1 as a floor, never a guarantee.
Sorted by price rather than by fame. The reference column is the number of images the model's endpoint will actually accept, which is often the thing that decides the choice.
| Model | Wallet | One picture | Reference images | Opens at |
|---|---|---|---|---|
| GPT Image 2 | neither | 0 credits | up to 6 | Free, 3 runs a day |
| Z-Image Turbo | ⚡ green | from ⚡1 | none | Free |
| Qwen Image Edit (Local) | ⚡ green | from ⚡1 | up to 3 | Free |
| Krea 2 Identity | ⚡ green | from ⚡1 | 1 required, 2 max | Free |
| Krea 2 Turbo | ⚡ green | from ⚡1 | none | Free |
| SeedVR Upscale | ⚡ green | from ⚡2 | 1 | Starter |
| Qwen Image Edit | ✦ gold | ✦1.4 | up to 3 | Starter |
| Seedream 4.5 | ✦ gold | ✦1.5 | up to 6 | Starter |
| Qwen Image 3.0 | ✦ gold | ✦1.5 | up to 3 | Starter |
| Qwen Image 3.0 Pro | ✦ gold | ✦1.8, ✦3 at 2K | up to 3 | Starter |
| Nano Banana | ✦ gold | ✦2 | up to 6 | Starter |
| Seedream 5.0 Pro | ✦ gold | ✦2.2, ✦3.7 at 2K | up to 6 | Starter |
| Nano Banana Pro | ✦ gold | ✦4.7, ✦8.1 at 4K | up to 6 | Starter |
| Qwen Image Edit 2.0 Pro | ✦ gold | ✦7 | 1 to 3 | Starter |
| Qwen Image Edit Max | ✦ gold | ✦7 | 1 to 6 | Starter |
Three of those rows deserve a sentence. Nano Banana Pro is the only image model here whose provider price depends on the output size, so 1K and 2K cost the same ✦4.7 and 4K costs ✦8.1 — the same ratio the provider charges us. Seedream 5.0 Pro and Qwen Image 3.0 Pro each carry a two-step resolution price instead, and both are quoted at the worst case their picker can produce, because the composer has to name one number before it knows how many references you are going to attach. SeedVR Upscale runs on our GPU in green, but if that lane is unavailable the same job is served over the paid API at ✦4.
Note: GPT Image 2 costs 0 credits on every plan and takes up to six references, with a daily allowance that rises from three runs on Free to two hundred on Ultra. It rides a single shared subscription rather than a metered key, which is also its limit: when that subscription's daily allowance is exhausted, the model pauses for everyone until it resets. It is a real option, not a way to avoid spending.
Open any result and there is a second, separate price list. These are one-click transformations on an existing image, each a hosted call with its own rate, and every one of them is billed in gold. There is no tier gate on them — but there is no green route either, so in practice they are a paying account's feature.
| Tool | One run |
|---|---|
| Upscale, and Upscale · Crisp | ✦0.5 |
| Upscale · Creative | ✦1 |
| Enhance · Restore | ✦1 |
| Relight | ✦1 |
| Inpaint | ✦1 |
| Angles | ✦1 |
| Enhance | ✦1.5 |
| Enhance · Sharpen | ✦2 |
| Relight · Custom | ✦2.5 |
Gold credits arrive two ways: monthly with a plan, or bought as one of five bundles priced on your own tier's ladder. Buying a bundle never changes your tier, so it opens no model that your plan does not already open. The cheapest rate on the site is Ultra's annual grant, and the whole catalogue's margin floor is set against it.
| Plan | Per month, billed annually | Gold a month | Cost of one credit |
|---|---|---|---|
| Free | — | none | — |
| Starter | $5.00 | ✦50 | $0.1000 |
| Creator | $34.99 | ✦380 | $0.0921 |
| Pro | $75.00 | ✦900 | $0.0833 |
| Ultra | $750.00 | ✦10,000 | $0.0750 |
So a ✦4.7 Nano Banana Pro picture is somewhere between 35 and 47 cents depending on which plan's grant paid for it, and a ✦1.4 Qwen Image Edit is between 11 and 14 cents. A picture on our own GPU is not on this table at all, because green credits are not for sale at any price — they arrive with the plan and refill on a clock.
We are not going to quote anybody else's pricing page, because we do not bill their models and cannot check the number. What travels between platforms is the set of questions, and they are worth asking of us as well: at what output size is that price quoted, does it change at 2K and 4K, does attaching reference images surcharge it, is the allowance a one-off block or something that refills, and does the cheapest tier reach the model that was in the headline. Four of those five have moved a price by more than 2× somewhere in the table above.
Open the image composer — The quoted cost sits on the Generate button before you press it. Nothing runs until you do.
GPT Image 2 at 0 credits, within its daily allowance and while its shared capacity holds. After that, the own-GPU models — Z-Image Turbo, Qwen Image Edit, Krea 2 Identity and Krea 2 Turbo — from ⚡1 a picture out of a wallet that refills toward ⚡20 a day on the free plan.
About twenty on the own-GPU lane, where pictures start at ⚡1. It is deliberately "about": own-GPU rates are set at runtime and can carry a demand multiplier when the queue is deep, so twenty credits is not a promise of twenty pictures.
No. Those are hosted models billed to us per call, so they are paid in gold and open at the Starter plan. The picker shows them with the plan that opens them rather than hiding them.
Because its provider charges more at 4K, by a ratio we pass straight through. It is the only image model here whose provider price moves with the output size — every other row is flat at any size its picker offers.
No. Bundles buy gold, which pays for hosted models and queue priority. They do not change your tier, do not open a gated model, and do not raise the daily green allowance — only the plan does that.
In the lightbox, ✦0.5 for the standard and crisp upscalers and ✦1 for the creative one, all in gold. SeedVR Upscale in the composer runs on our own GPU from ⚡2, and opens at Starter.