Twilight edit for 20 listing photos with GPT Image 2.5: what it costs

A twilight edit of 20 listing photos with ChatGPT Image 2.5 on Sume costs about $0.27 at medium quality and $1.03 at high, using one reference image per call at 1536x864. Those are estimates from the repository pricing code, with input image tokens estimated, so read the live row with GET /v1/images/models before you quote a client.

The pattern is one edit call per photo: the original photo goes in as the single reference, the prompt asks for dusk light and lit windows, and aspect_ratio: "auto" keeps the frame shape.

The request

Reference URLs must be public HTTPS, and the model reads the first image as the one to edit. aspect_ratio: "auto" matches the reference on edit calls, and the docs warn that omitting the field is not the same as sending auto.

{
  "model": "openai/gpt-image-2.5",
  "prompt": "Same house and framing, dusk sky, warm lit windows, no new objects",
  "quality": "medium",
  "aspect_ratio": "auto",
  "input_references": [
    {"type": "image_url", "image_url": {"url": "https://example.com/listing-01.jpg"}}
  ]
}

Cost for 20 photos

The per-call numbers below include one estimated reference image. The no-reference column is the plain text-to-image estimate at the same size, for comparison.

A review step that pays for itself

A listing photo edit has one failure that costs far more than the image: a change the buyer would call misleading. Add a short human check to the workflow. Put the original and the edit side by side, confirm that the building, the grounds and the neighbors are unchanged, and only then export.

Because a rejected edit costs only a few cents, it is cheaper to regenerate than to fix. If one window looks wrong, rerun that photo at the same quality and compare. For a stubborn case, switch to a masked edit so the house stays pixel-identical and only the sky is touched; the mask is an optional mask_url that must be public HTTPS.

• Keep the original files untouched and store the edit separately.
• Label edited photos where a portal or local rule requires it.
• Log the prompt and quality used for each listing.

What to watch

A real-estate edit has one hard rule: do not add or remove anything that changes what the buyer is buying. Keep the prompt to lighting and sky, and compare the output with the original before it goes on a listing. Rules for disclosure of edited listing photos differ by portal and region, and this post does not cover them.

Each edit call waits up to 30 seconds in sync mode, and slower configurations such as high quality can return a 202 job envelope instead. Read the status code, not the body shape, as the Image API docs say.

For a larger run, the same request works in a loop. Twenty listings at about a cent and a third each is a small number, but a brokerage with 400 listings a month would see it multiply, which is why the quality tier is the first lever to check. Keep a table of listings, the quality used, the cost reported in usage.cost, and whether the photo was approved. After a month, the data tells you whether medium was enough, and you can stop paying for high on every photo.

• Use quality: medium for drafts and high for the photos you publish.
• Send one reference per call; extra references raise the input token estimate.
• If only the sky should change, add a mask_url over the sky region.