GPT Image 2.5 at 3840x2160: $0.1251 at high, by quality

A 3840x2160 image from GPT Image 2.5 on Sume costs about $0.0140 at low quality, $0.0325 at medium, $0.1251 at high, $0.2224 at xhigh and $0.5004 at max, by my estimate for a text-only prompt. That is the largest 16:9 size the model takes, since 3840 by 2160 is exactly 8,294,400 pixels. At high, it costs less than twice a 1024x1024 image.

All figures assume a text prompt with no reference images and no more than one image per call.

Where 3840x2160 sits in the rules

OpenAI's image guide, read on 2026-10-07, sets the limits that Sume's docs repeat: both edges multiples of 16, a long edge up to 3840, a ratio up to 3:1, and 655,360 to 8,294,400 pixels in total. 2160 divided by 16 is 135, so 3840x2160 is valid, and its pixel count equals the maximum. Any taller 4K, such as 3840x2176, is over the limit.

Whether you need 4K at all is a separate question. A 3840x2160 file suits a large display or print; for a web page, a 1536x864 or 2560x1440 image is usually enough. Check where the image will be shown before paying for the top row.

The table

Each figure is the provider list price times 1.25, before cent rounding, from the output-token formula at $30 per million tokens. Prices at the same quality do not rise in step with pixel count, because the formula also depends on aspect ratio; by the formula, 1536x864 comes out below 1024x1024 at every tier. Treat the table as an estimate and read usage.cost for the amount billed.

The 2560x1440 column suits most screens: at medium it is about $0.0180, a little over half of the 4K price at the same tier.

Choosing a tier

For a large format, the tier matters more than the size. Going from 1536x864 to 3840x2160 at high costs about 3.1 times as much ($0.1251 divided by $0.0405), while going from medium to max at 3840x2160 costs about 15.4 times as much ($0.5004 divided by $0.0325). Draft at low or medium, and render the final at high.

Slow settings may come back as a 202 job envelope rather than a 200. Sume's docs list 4K, high quality and large n as the likeliest to run longer than the 30-second wait, so branch on the status code.

For a batch, multiply by the count and the n value. Ten 4K images at high are about 10 x $0.1251 = $1.25 before rounding, while ten at medium are 10 x $0.0325 = $0.325. Round each call up to the cent when you budget, since a per-call amount of $0.0325 cannot be billed as a fraction of a cent.

A 4K request

The request pins every setting that moves the price. The key comes from the environment, and the size is explicit.

The response includes usage.cost, which is the billed USD amount for that call. Log it for the first few runs and compare with the table above; if your numbers differ, trust the log.

curl -sS -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": "A panoramic city skyline at dusk, crisp detail",
    "quality": "high",
    "image_size": {"width": 3840, "height": 2160}
  }' -w '\nHTTP %{http_code}\n'