GPT Image resolution above 2560x1440: experimental, and Sume's cap

OpenAI's image guide says resolutions above 2560x1440 are experimental for GPT Image models. Sume's ChatGPT Image 2.5 models accept custom pixels up to a 3840 edge, so a 3840x2160 request is valid on the Sume side, but treat anything above 2560x1440 as the experimental zone OpenAI describes.

Sume rules are from the Image API docs; the experimental line is from OpenAI's guide. Both were read 2026-09-30. The Sume docs do not say whether sizes above 2560x1440 are experimental on Sume.

What are Sume's custom pixel rules?

For the 2.5 models, image_size accepts named presets, auto, or custom pixels. Custom pixels need both edges to be multiples of 16, a maximum edge of 3840, an aspect ratio of at most 3:1, and a total of 655,360 to 8,294,400 pixels. OpenAI's guide lists the same limits and adds that the 8,294,400 ceiling is 4K.

Where do I put the pixels?

In image_size, not size. The docs call size a shorthand for a resolution tier and say not to put custom pixels on it; the parameter table adds that explicit pixels are not served on size in v1. The separate resolution field takes a tier (512, 1K, 2K, 4K).

const response = await fetch("https://api.sume.com/v1/images", {
  method: "POST",
  headers: {
    Authorization: "Bearer " + process.env.SUME_API_KEY,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    model: "openai/gpt-image-2.5",
    prompt: "A wide product shot on a pale studio floor",
    image_size: "2560x1440",
  }),
});
console.log(await response.json());

What happens if the size breaks a rule?

The docs state that a request that sets a parameter the selected model does not list is rejected with 400 unsupported_parameter. That sentence is about unlisted parameters; the docs do not spell out the error for an out-of-range value. Validate sizes in your own code first rather than depending on a particular error.

Should I go above 2560x1440 at all?

If you need a 4K deliverable, you have two routes: ask for a 4K-class size directly, accepting OpenAI's experimental label, or generate at or under 2560x1440 and enlarge afterwards. The second route is covered in AI image 4K: generate or upscale. Pick by testing a few of your own prompts; this page does not rank them.