Per-image rates for GPT Image 2 and the Gemini image models at 1K, 2K and 4K, from the cost table atelier.space runs on — including the size tier that is free to upgrade.
Updated 2026-07-25 · figures verified 2026-04-28
Image models are quoted in tokens, in credits, or in nothing at all. The number that answers "what will this cost me" is dollars per finished image at a given size. This page is that number, from atelier.space's own vendor-cost table — which records the providers' published rates and is used for quota burn-down rather than as a billing substrate.
| Model | 1K | 2K | 4K |
|---|---|---|---|
| Nano Banana 2 | $0.067 | $0.101 | $0.151 |
| Nano Banana Pro | $0.134 | $0.134 | $0.240 |
| GPT Image 2 | $0.053 | $0.150 | $0.320 |
Source: server/cost/pricing.ts, PRICING.image.models. Checked against the vendor pricing pages on 2026-04-28.
On Nano Banana Pro, 1K and 2K cost the same. Google bills that model per output token and fixes the count at 1,120 tokens for both tiers, so the larger image is a free quality upgrade. There is no reason to ask it for 1K.
GPT Image 2 does not behave that way — its rate scales with size, and it also has quality tiers underneath the size. At 1K, low quality is $0.006 and high is $0.211, a spread of roughly 35×. If you are iterating on composition, iterate low and finish high.
Source: server/cost/pricing.ts, the per_quality block on gpt-image-2.
The rate table registers more models than any single deployment serves. On atelier.space's production lanes, the live image models are Seedream 5.0 Pro and GPT Image 2; the Gemini rows are registered for deployments that route through Vertex or LiteLLM and render as not-configured otherwise.
Source: src/nodes/imageSettings.ts, the deployment note above the registry.
Seedream is billed per output image by its platform rather than by a size ladder, which is why it does not appear in the table above; its cost sits in the same file under the Ark lane.
GPT Image 2 will not accept an arbitrary size. The rule derived in src/nodes/imageSettings.ts is: the long edge must be 3840 or less, both edges must be multiples of 16, the aspect ratio must not exceed 3:1, and the total pixel count must fall between 655,360 and 8,294,400. Sizes outside that are rejected at the vendor's door rather than clamped.
Seedream does not offer 8:1 for a related reason — at that ratio the short edge falls below 512 pixels.
response_cost, atelier's quota layer prefers the live figure over the static one here, so a small drift shows up on quota accounting rather than on what you are charged.