Sume Auto or a pinned model for repeatable image edits?

OpenAI's Images 2.5 thread says ChatGPT automatically selects the Flare variant for routine prompts and escalates to Sunburst for complex requests that need precise text or intricate layouts (OpenAI Developer Community, read 2026-10-04). Routing like that is convenient inside a chat app. For an edit you must be able to repeat next month, convenience is the problem.

What Sume Auto does and does not tell you

The Image API docs describe model: "sume/auto" as the way to let the Image Router choose. The response echoes the id you requested, so sume/auto stays sume/auto, and the job model stays sume/auto. The catalog never lists Auto as a row. Sume does not disclose which family served an Auto request.

When to use which

Sume lists openai/gpt-image-2.5 and openai/gpt-image-2.5-sunburst as separate ids, so a pinned request also picks the variant. Choose Sunburst only if you tested that it helps on your layouts.

• Exploring, with no need to repeat the result: Auto is fine, and on MCP omitting the model routes to it.
• A brand look you will repeat, an A/B test or a cost comparison: pin the id.
• A support complaint about one image: you can only answer it if you pinned, or logged, the request.

Pin it in the request

Send this to POST /v1/images. Log the model id with the result so the pin is on record.

{
  "model": "openai/gpt-image-2.5",
  "prompt": "Swap the headline to \"Autumn Sale\" and keep the layout",
  "input_references": [
    {"type": "image_url", "image_url": {"url": "https://example.com/layout.png"}}
  ],
  "aspect_ratio": "auto"
}