GPT Image 2.5 API pricing: what one image costs
See current GPT Image 2.5 API pricing per image for Flare and Sunburst, compare flat and token billing, and estimate a monthly budget from real rates.
Read as MarkdownGPT Image 2.5 API pricing on SeedRouter comes in two forms. Standard model IDs charge one flat price for every delivered image, whatever the size or quality. Official model IDs bill the tokens each render reports, so the price moves with quality and size. Flare and Sunburst cost the same within a channel. Failed requests are not charged.
The current numbers are below, read live from the rate table that bills your account. Nothing on this page is a remembered quote.
What does one GPT Image 2.5 image cost right now?
Current rates are temporarily unavailable. No price estimate is provided; unavailable rates must not be interpreted as free usage.
Read the table in two parts. The first rows are the rate itself: a price per delivered image on Standard, and a price per million input and output tokens on Official. The second table converts both into the cost of one 1024×1024 image at each quality step, so the two channels can be compared on the same line.
If the rates cannot be loaded, the table says so instead of showing zero. Treat an unavailable rate as unknown, never as free.
Standard or Official: which billing fits your workload?
The choice is about predictability versus paying for exactly what a render uses.
| Workload | Better fit | Why |
|---|---|---|
High quality steps (high, xhigh, max) | Standard | The flat price does not rise with the quality step |
| Large canvases, up to 3840 pixels on an edge | Standard | Size does not change a flat price |
Many small drafts at low | Official | Low-quality renders report few output tokens |
| A fixed budget per image, known in advance | Standard | The cost of a request is the image count times one price |
Both channels accept the same 12 request fields and return the same task response, so the decision can be made per request. A common pattern is to draft on the channel that is cheaper for small renders and produce finals on the one that is cheaper for large ones.
How do you estimate a GPT Image 2.5 request?
On a Standard ID the estimate is exact before you submit:
cost = delivered images × price per imageAsk for n: 4 and receive four images, and the charge is four times the per-image price. Receive three, and you pay for three.
On an Official ID the estimate depends on token usage:
estimated USD = (
input tokens × input USD per million tokens
+ output tokens × output USD per million tokens
) / 1,000,000Output tokens depend mostly on the quality step and the canvas. Measured at 1024×1024, one render reported 196 output tokens at low, 439 at medium, 1,756 at high, 3,122 at xhigh and 7,024 at max. Those are single observed samples, not guarantees. A different subject or a larger size lands near them, not on them. The pricing calculator on the model page applies the same numbers to the size, quality and image count you choose.
What changes the bill, and what does not?
On Standard IDs only the number of delivered images changes the bill. Prompt length, reference images, quality and size do not.
On Official IDs several settings move the cost:
- Quality step. Each step up spends more output tokens;
maxmeasured about 36 timeslowat the same size. - Canvas size. A larger
sizeproduces more output tokens than a smaller one at the same quality. - Image count.
nmultiplies the output side. Ten images are ten renders, not one. - Prompt and references. These are input tokens, usually a small share of the total next to image output.
The model choice is not on either list. Flare and Sunburst are priced identically within a channel, so pick between them on speed and edit control, as described in Flare vs Sunburst.
How much should you budget per month?
Start from finished images, not requests. If a project needs 200 final images a month and a usable image takes three attempts, the budget covers 600 renders. Multiply by the Standard price for a fixed number, or by the Official estimate at the quality step you draft and finish with.
Then measure. Run a week of real work, compare the final charges in usage history with your estimate, and adjust the attempts-per-image figure. Credits are bought as one-time top-ups and do not expire, so an unused balance carries over to the next month.
When is a request not charged?
A task that ends failed is not charged, and neither is a request that returns no image. If you ask for several images and fewer come back, you pay for the ones delivered.
A polling timeout on your side is not a failure. Keep the task ID and check the same task again before submitting a replacement; submitting twice can mean paying twice. The task guide explains the difference.
Frequently asked questions
Is there a free tier for the GPT Image 2.5 API?
No. SeedRouter bills prepaid credits, and every delivered image is charged at the current rate. Failed requests cost nothing, which keeps experiments with new prompts cheap.
Do Flare and Sunburst have different API costs?
Not on SeedRouter. Within a channel both models have the same price. The channel, Standard or Official, decides how you are billed.
Where can I see what a request actually cost?
In account usage history. The task response reports token usage when it is available, but it carries no dollar amount.
Price the request before you send it
Decide the billing channel per job: a flat price when quality or size is high, token billing when you draft small. Check the live table above or the GPT Image 2.5 model page before a large run, and use the API reference for the exact fields.
