GPT Image 2.5 above 2560x1440 is experimental: a safe size ladder

OpenAI's image generation guide says custom sizes for the GPT Image models must have both edges as multiples of 16, an aspect ratio between 1:3 and 3:1, a maximum edge of 3840 pixels and between 655,360 and 8,294,400 total pixels. It also says resolutions above 2560x1440 are experimental.

Sume's Image API page states the same rules for image_size on ChatGPT Image 2.5: edges in multiples of 16, maximum edge 3840, aspect at most 3:1, 655,360 to 8,294,400 pixels. Custom pixels go on image_size as {width, height}; the size field takes tier shorthand only and rejects WxH.

A ladder you can reuse

These sizes pass the multiple-of-16 and pixel-range rules. Only the last row sits above the experimental line; 2560x1440 is the top of the stable range.

Validate before you send

A malformed size costs a round trip. This check mirrors the documented rules:

def valid_gpt_size(w: int, h: int) -> bool:
    if w % 16 or h % 16:
        return False
    if max(w, h) > 3840:
        return False
    if max(w, h) / min(w, h) > 3:
        return False
    return 655_360 <= w * h <= 8_294_400

for size in [(2560, 1440), (3840, 2160), (4096, 2160), (1000, 1000)]:
    print(size, valid_gpt_size(*size))

Plan for the experimental tier

Treat anything above 2560x1440 as a draft you inspect, not a guaranteed output. Large, high-quality requests can also run past the 30 second wait on POST /v1/images; Sume then returns a 202 with a job to poll rather than an error. If you need 4K stills for print, generate at 2560x1440 first and compare against a 3840x2160 run on the same prompt.

How this was checked

Vendor facts come from the pages listed in the sources, read on 2026-10-05. Sume facts come from the Image API docs and the catalog code on main on the same date. Catalogs and limits change, so read the descriptors from GET /v1/images/models before you pin a number in production code.