2026-10-08

GPT Image 2.5 draft grid: four low-quality images for 4 cents

Four GPT Image 2.5 drafts in one Sume call at quality low cost 4 cents; a dollar buys 25 such calls. The arithmetic, the n field and when to go to high.

5 min readSume

All posts

A draft grid of four GPT Image 2.5 images at quality: "low" costs 4 cents on Sume: 4 images times 1 cent each. One dollar therefore buys 25 four-image calls, or 100 draft images. Send n: 4 and quality: "low" to POST /v1/images, pick the layout you like, then re-render only that prompt at high.

What the request looks like

ChatGPT Images 2.5 is on Sume as openai/gpt-image-2.5 (Flare) and openai/gpt-image-2.5-sunburst. The Image API docs say that if you omit quality, the default is high, so a draft grid needs the field set explicitly. OpenAI's own guide says to use quality: "low" for quick drafts and to compare higher settings for final assets.

The docs also say Sume bills cost_usd x n, and that n goes up to 10 per call with lower per-model ceilings. The catalog row for this model lists a maximum of 4 images per call.

{
  "model": "openai/gpt-image-2.5",
  "prompt": "flat-lay of a ceramic mug on linen, soft window light",
  "quality": "low",
  "n": 4,
  "aspect_ratio": "4:5"
}

Draft cost against final cost

Billable per image on Sume, GPT Image 2.5 at 1024-class sizes (as of 2026-10-08, from the Sume price catalog)
Quality	Billable per image	Four images	Calls per $1 at n=4
low	1 cent	4 cents	25
medium	2 cents	8 cents	12
high	7 cents	28 cents	3

Why draft low first

The gap between low and high is 6 cents per image, so a grid of four drafts costs less than one high render. A common pattern is two draft rounds (8 cents), then one high render of the winning prompt (7 cents): 15 cents for a result you chose from eight options, against 56 cents for eight high renders.

Low quality is for composition, not for fine text. If the final needs a legible headline, check the text at high quality rather than trusting the draft; OpenAI notes that text placement and clarity can still struggle.

Set quality on every draft call; the omitted default is high.
Keep the same prompt and aspect_ratio between draft and final so the layout carries over.
Expect different pixels at high quality: the model does not take a seed on Sume, so a re-render is a new image, not a sharper copy of the draft.

A worked example: a mug listing

Say you need a hero image for a ceramic mug listing and you are unsure about the setting: linen, oak table or a windowsill. Run the prompt three times at n: 4, quality: "low", once per setting. That is 12 drafts for 12 cents. You pick one, then re-render it once at high for 7 cents. The whole exploration costs 19 cents, less than three high-quality renders made blind.

The same pattern works for thumbnails and social cards, where the choice you are making is composition and crop, not pixel detail. For the other direction, see how the price climbs with xhigh and max in the xhigh and max price grid.

Limits to keep in mind

n is bounded by the model row, so a request above 4 for this model is not a way to get a bigger grid; send two calls instead. A call that does not finish inside the 30-second wait returns 202 with a job envelope, and you then read the images from the job result endpoint. Low quality is the least likely setting to hit that, since the docs name 4K, high quality and large n as the slow configurations.

Failed or cancelled generations are not charged, so a retry after an error does not double the bill.

Rounding note

List price for the model in the catalog is $0.0527 at high 1024 output. Times 1.25 that is $0.065875, which the catalog rounds up to 7 cents. Check usage.cost in the response for the amount billed to your wallet on each call.