ChatGPT Images 2.5 Sketch: the API way is a drawing as a reference

If you want ChatGPT Images 2.5's Sketch behavior through an API, there is no sketch canvas on Sume, but you can get the same result by uploading your drawing to a public https address and sending it in input_references to openai/gpt-image-2.5 with a prompt that says what the drawing is. The model starts from your layout instead of from nothing.

9to5Mac reports that OpenAI released ChatGPT Images 2.5 on 2026-09-08, that a Sketch feature lets users draw directly in ChatGPT as reference material, and that two API models exist: GPT-Image-2.5 Flare for general use and GPT-Image-2.5 Sunburst for premium creative work (read 2026-10-06, 9to5Mac). The drawing tool is a ChatGPT product feature. Sume ships the models, not that interface.

What Sume supports

Per the Sume Image API docs, both openai/gpt-image-2.5 and openai/gpt-image-2.5-sunburst support text-to-image, up to 16 image references, an optional mask_url and a background setting. Reference URLs must be public https; Sume rejects localhost, private-network and non-https URLs before it submits the job.

The call

Describe the drawing in the prompt, then say what to make of it. The model uses the reference as the source, so a clear sentence such as this is a floor plan, render it as a bright room works better than a bare instruction.

import os
import requests

body = {
    "model": "openai/gpt-image-2.5",
    "prompt": (
        "The reference is a hand sketch of a poster layout: a big circle "
        "at the top, a headline band, two columns below. Render it as a "
        "finished poster with a sunrise theme, keeping the layout."
    ),
    "input_references": [
        {"type": "image_url", "image_url": {"url": os.environ["SKETCH_URL"]}}
    ],
    "quality": "medium",
}
r = requests.post(
    "https://api.sume.com/v1/images",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    json=body,
    timeout=60,
)
print(r.status_code)
print(r.json()["data"][0]["url"] if r.status_code == 200 else r.text[:300])

Flare or Sunburst for a sketch

9to5Mac describes Flare as the general-purpose model and Sunburst as the premium option for tighter editing control. For a first pass from a rough drawing, Flare is the natural start, and Sunburst is something to try on a final that needs closer fidelity to your lines. Compare the same sketch on both and keep the cheaper one if you cannot tell the difference. The Sume docs say both use the same token rates, so the choice is about results, not price; confirm in the pricing lines.

The drawing guides layout and does not pin pixels. If the model moves a line you care about, add a sentence naming it, or use a mask edit for a precise change.

Tips for a drawing the model can use

The quality of the result depends on the drawing, and a few habits help even a rough sketch.

• Draw big, dark lines on a plain light background and photograph or export them flat.
• Label parts with words only if you want those words in the picture.
• Say in the prompt what the drawing is, such as a floor plan or a poster layout.
• Keep one idea per sketch; split a busy page into several calls.
• Host the file at a public https address; local files cannot be fetched.

Cost and speed expectations

Do not assume a reference makes a call cheaper or faster. The Sume docs describe GPT Image 2.5's token rates, and the billed amount comes from the endpoint pricing line, so read that line for the model you use rather than assuming a sketch is free. OpenAI's launch material, as summarized by 9to5Mac, claims latency up to 50 percent lower than Images 2.0. Treat that as a vendor claim for ChatGPT, and time your own calls on Sume, because a high quality render can run past the 30-second wait and return a job envelope instead of the image.

What to do next

Once you have a render you like, treat further changes as edits to that image rather than new sketches; see multi-round edits on GPT Image 2.5. A related route that works on a photograph is drawing boxes on a photo as markup.