GPT Image 2.5 sizes 1024x1024, 1536x1024, 1024x1536: Sume price

OpenAI's image generation guide (read 2026-10-05) recommends three sizes for GPT Image 2.5: 1024x1024, 1536x1024 and 1024x1536. On Sume the two non-square sizes are cheaper than the square one at every quality: at high, 1536x1024 and 1024x1536 each quote $0.0515 against $0.0659 for 1024x1024, which is 22% less.

The price follows the output token count, and Sume's Image API page says output estimates use OpenAI's size and quality calculator with $30 per million output image tokens. Landscape and portrait have the same pixel count, so they quote the same amount. The prices below apply to both Flare and Sunburst.

Price by size and quality

The table adds 2560x1440 and 3840x2160 so you can see where larger canvases land. Each figure is Sume's quote with a one-character prompt.

Why a bigger image can cost less

Output tokens come from a grid, and the grid depends on the shorter side relative to the longer one for each quality. A 1536x1024 image has a 2:3 shape, and the calculator gives it fewer tokens than a square even though it has 50% more pixels. The result is counter-intuitive, so do not infer cost from pixel count alone.

The upper end behaves differently. At high, 3840x2160 quotes $0.1251, about 1.9x the square price, and it sits exactly on the 8,294,400-pixel cap that Sume documents for custom sizes.

A worked example: 500 landscape banners

Take 500 banners at high. At 1024x1024 the quote is $0.0659 each, so the batch is $32.94. At 1536x1024 it is $0.0515 each, or $25.75, a difference of $7.19 for a shape that is also closer to a banner. The saving is 19% at low, 22% at medium and 22% at high, because the token grid scales with the quality tier.

The saving is smallest in cents at low, where a 1536x1024 image quotes $0.0060 against $0.0074, a gap of a fraction of a cent. Do the arithmetic for your own volume before you change a pipeline for it.

Using a size in a request

Pass the size as image_size. For custom pixels both edges must be multiples of 16, the longest edge at most 3840, the aspect ratio at most 3:1, and the pixel total between 655,360 and 8,294,400. Named presets and auto are also accepted; auto and presets without a verified pixel mapping reserve the upper bound while the job runs.

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-sunburst",
    "prompt": "Wide product shot of a linen jacket on a hanger",
    "quality": "high",
    "image_size": "1536x1024"
  }'

When to pick which

Pick the shape the image will be shown in, not the cheapest one: a banner cropped from a square wastes the pixels you paid for. If the target is a feed post, a 1536x1024 or 1024x1536 render is both closer to the final shape and cheaper than a square.

Check the quote before a large batch. Send one call at the size you plan to use, read usage.cost from the response, and multiply by the count; a different size will not match this table.

One caveat on the cheap non-square sizes: they only help if the shape fits. A 3:2 image shown in a 4:5 slot is cropped, and the crop throws away pixels you paid for. For feed formats that are taller than 2:3, such as 4:5, read the aspect ratio breakdown before you decide.

Choosing a size

Pick the frame from where the image will be used, not from the price. A landscape banner wants 1536x1024, a portrait card wants 1024x1536, and a square thumbnail wants 1024x1024. Because the two rectangles are cheaper than the square at every tier in the table, the price never argues for forcing a square.

If you need a frame that is not one of the three, send custom pixels that follow the multiple-of-16 rule, and read the first response's usage.cost against the table. A size you did not test is the one most likely to surprise a budget, and the check costs a fraction of a cent at low.