GPT Image 2.5 API on Sume: quality, sizes, masks, and references

To use the GPT Image 2.5 API on Sume, send POST /v1/images with model: "openai/gpt-image-2.5" (ChatGPT Image 2.5, which the docs call Flare) or "openai/gpt-image-2.5-sunburst", plus a prompt. Both ids generate from text or edit from up to 16 reference images, with a quality scale up to max, custom pixel sizes, an optional mask_url, and a background setting.

Everything here comes from the Image API docs and the catalog Sume serves at GET /v1/images/models, read 2026-09-26. ChatGPT Image 2 (openai/gpt-image-2) remains selectable, with a shorter option list.

What is the difference between Flare and Sunburst?

The docs and the catalog describe no difference in what you can send. Both ids support text-to-image, up to 16 image references, an optional mask_url, and background: auto|transparent|opaque. The catalog names them ChatGPT Image 2.5 and ChatGPT Image 2.5 Sunburst, gives them identical parameter descriptors, and prices both from the same token rates. The docs do not say how their images differ, so compare both on your own prompts before you pin one.

Which quality levels can I set?

Both 2.5 ids accept auto, low, medium, high, xhigh, and max for quality; omit the field and the request runs at high. ChatGPT Image 2 takes only low, medium, and high.

The level feeds the token estimate, so it changes the price. High settings are also among the slow configurations that can return a 202 job envelope instead of an image; 4K, quality, and image count covers that case.

Which sizes does ChatGPT Image 2.5 accept?

Set image_size to a named preset, auto, or custom pixels as { width, height } or WIDTHxHEIGHT. A custom size needs both edges in multiples of 16, a longest edge of at most 3840, an aspect ratio no wider than 3:1, and 655,360–8,294,400 pixels in total. 3840×2160 sits exactly on the edge and pixel ceilings.

aspect_ratio works too: the three ChatGPT Image ids list auto, 1:1, 16:9, 9:16, 4:3, 3:4, 5:4, 9:8, 4:5. Ratios without a named preset become legal custom pixels. 4:5 turns into 1024×1280, because 1080×1350 fails the multiple-of-16 rule; 5:4 turns into 1280×1024 and 9:8 into 864×768. Other families are compared in aspect ratios and custom sizes.

How do I edit an image with a mask?

Put the source image in input_references, a public HTTPS mask image in mask_url, and prefer aspect_ratio: "auto" so the output follows the source. Only the 2.5 ids list mask_url and background; ChatGPT Image 2 answers either field with 400 unsupported_parameter. General reference rules are in image generation with reference images.

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": "replace the sky with a soft sunset",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/photo.jpg" } }
    ],
    "mask_url": "https://example.com/mask.png",
    "quality": "medium",
    "aspect_ratio": "auto"
  }'

How do ChatGPT Image 2.5 and ChatGPT Image 2 compare?

They share the custom-size box, the ratio list, and the named presets. The 2.5 ids add auto, xhigh, and max quality, six more reference slots, masks, and background.

How is GPT Image 2.5 billed?

By tokens: output image tokens, input image tokens, and input text tokens, at the same rates for Flare and Sunburst. The catalog pricing line for both ids is a single figure for high quality at 1024×1024 output. Each real request is estimated from its size and quality, and its input token counts are estimates.

When Sume reserves funds, quality: "auto" reserves as max and auto size reserves the output-token upper bound. The response's usage.cost is the USD amount billed to your wallet, and a failed or cancelled generation is not billed. Catalog lines include Sume's margin, plus a 5.5% agent fee by default; see API pricing.