GPT Image 2.5 largest square: 2880x2880, not 3000x3000

The largest square GPT Image 2.5 accepts as a custom size is 2880x2880. That is exactly 8,294,400 pixels, the documented cap. 3000x3000 is 9,000,000 pixels and fails the pixel rule even though it is under the 3840 edge limit.

The rules are from OpenAI's image guide and the Sume Image API docs; both read 2026-10-01. The square arithmetic is ours.

What are the limits?

OpenAI's guide lists the GPT Image 2.5 custom-size rules: both edges multiples of 16, edges at most 3840, aspect ratio between 1:3 and 3:1, and 655,360 to 8,294,400 total pixels. Sume's Image API docs state the same limits for image_size on GPT models.

Is 2880x2880 safe to depend on?

OpenAI's guide calls output above 2560x1440 experimental. 2880x2880 has 8,294,400 pixels versus 3,686,400 for 2560x1440, so it is in that experimental range. Test it before building on it.

How do I request it?

Send the pair as a string.

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": "Seamless watercolor pattern of ferns, soft green on cream",
    "image_size": "2880x2880",
    "output_format": "png"
  }'

What if I need more pixels?

Generate within the cap and upscale. Sume's Image Upscale 1.0 takes upscale_factor from 1 to 4 and costs $0.20 per image. A 4x pass on a 1024x1024 image is 4096x4096 by plain multiplication; the docs do not promise detail beyond that.