GPT Image 2.5 widest banner: 3072x1024 at 3:1, with output cost

GPT Image 2.5's widest legal shape is 3:1, and 3072x1024 fits every rule: both edges are multiples of 16, neither edge exceeds 3840, and the area is 3,145,728 pixels, inside the 655,360 to 8,294,400 range. At that size the output estimate is $0.00741 on medium and $0.02964 on high. A 4:1 or 8:1 banner is outside the model's limits, so use a model that lists those ratios.

The rules, applied

OpenAI's image generation guide gives the custom size limits, and Sume's Image API page repeats them: multiples of 16 on both edges, maximum edge 3840, aspect ratio at most 3:1, 655,360 to 8,294,400 pixels. 3072 is 192 x 16 and 1024 is 64 x 16. A 4096x1024 banner fails twice: the edge exceeds 3840 and the ratio is 4:1.

Output estimate at 3072x1024

Computed from the output token formula in Sume's pricing code for GPT Image 2.5 at $30 per million output tokens, before input tokens and Sume pricing. The docs themselves publish only the 1024x1024 xhigh and max figures. The values below are lower per pixel than the 1024x1024 figures because the short edge sets the token grid.

A request

Pass the size as image_size in WIDTHxHEIGHT form; image_size wins over aspect_ratio, and the docs say sizes without a verified GPT pixel mapping reserve the output upper bound, so an explicit size gives the tighter estimate. The normalized aspect_ratio list in the docs has no 3:1 value, so do not use it for this banner. OpenAI also notes that resolutions above 2560x1440 are experimental, so test 3072x1024 before you commit to it.

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": "A wide coastal panorama at dusk, empty sky on the left for headline text",
    "image_size": "3072x1024",
    "quality": "medium"
  }'

When you need wider

Taller ratios than 3:1 need another family: the Sume docs list 4:1, 1:4, 8:1 and 1:8 among the normalized aspect_ratio values, but a model only accepts the values its catalog descriptors list. Read supported_parameters first. A banner that is a multiple of 16 on one family may not be on another.