GPT Image 2.5 4K: how to request a 3840x2160 image by API

To get a 4K image from GPT Image 2.5, send POST https://api.sume.com/v1/images with model set to openai/gpt-image-2.5 and image_size set to 3840x2160. That size sits exactly on the model's edge and pixel ceilings, so it is the largest 16:9 output you can ask for. Expect a slow request: at this size the call may return 202 with a job to poll instead of the finished image.

Limits come from Sume's Image API docs, read 2026-09-29. OpenAI's guide notes that resolutions above 2560x1440 are experimental, so check every 4K result before you ship it.

What does the request look like?

Use image_size for a custom size such as 3840x2160; the docs describe it as the field for named presets, auto or custom pixels. The 2.5 ids advertise no resolution tier in the catalog. A parameter a model does not list is rejected with 400 unsupported_parameter, so read GET /v1/images/models first. mode: "async" skips the 30-second blocking wait and always returns the job envelope.

curl -X POST "https://api.sume.com/v1/images" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "aerial photo of a harbor town at sunrise, fishing boats, soft haze",
    "image_size": "3840x2160",
    "quality": "high",
    "mode": "async"
  }'

How do I get the image after a 202?

The 202 body has a job with status_url and result_url. Poll GET /v1/jobs/{id}/status until it is terminal, then read GET /v1/jobs/{id}/result for the images. Those are the standard job endpoints and return the standard job result shape, not the image body a 200 returns. You can take a webhook instead of polling: send mode: "webhook" with a public HTTPS webhook_url.

curl "https://api.sume.com/v1/jobs/$JOB_ID/status" \
  -H "Authorization: Bearer $SUME_API_KEY"

curl "https://api.sume.com/v1/jobs/$JOB_ID/result" \
  -H "Authorization: Bearer $SUME_API_KEY"

How much does a 4K image cost?

Should I generate 4K or upscale?

Both work; the choice is cost and control. Generating at 4K makes the model draw every pixel; upscaling a smaller image adds pixels to a picture you already approved. Price both for your own prompt: upscaling with the image API covers that path. Input tokens for any reference images are added to the figures above, and a failed or cancelled generation is not billed.