GPT Image 2.6 workflows: auto routing, references and chat channels

Open the image-model control beside the Agent composer options. Automatic lets the assistant route the request from its visual signals. A fixed choice names one of gpt-image-2.6, gpt-image-2.6-sunburst or gpt-image-2.6-flare. The preference is saved for your next new request, so you can keep the last stable choice without changing an already submitted task.

Automatic routing reads the task you are asking for, not just a single keyword. A plain subject or scene starts with the general model. People, multiple references, product identity and complex object relationships point toward sunburst. Posters, packaging, menus, diagrams and requests with a lot of in-image wording point toward flare. The route is a practical product policy and can fall back when a model is unavailable; it is not a provider quality ranking. If the choice matters, the task record keeps the resolved model so a retry does not silently change it.

For a series, a prompt test or a client review, choose one fixed model before the first image. Keep the aspect ratio, quality and reference set stable while changing one visual decision. This makes a comparison useful: you can tell whether the change came from the prompt or the model. If you need to change the model, make that a new attempt and label the result. Existing tasks retain their original model, quality, references and output settings.

A model choice cannot replace a clear brief. State the subject, purpose, audience, composition, aspect ratio, background, exact words and what the reference must preserve. For an identity reference, name the silhouette, material, label position and distinctive details. For a style reference, say which lighting or palette to borrow and which subject to ignore. This separation makes automatic routing easier and makes a fixed-model comparison fairer.

Telegram, Weixin, Lark and DingTalk can each keep an image-model preference. Set automatic for a general-purpose bot, or choose a fixed model for a channel dedicated to product cards, posters or a repeatable series. The setting affects new image tasks in that channel. A task already accepted by the channel keeps its frozen model even if an administrator changes the setting later. Check the channel’s result in Sparkles before sending a second request when delivery is uncertain.

Attach only the references that the request needs and identify their roles. Ask for PNG with a transparent background when you will composite a cutout, then inspect the alpha edges against contrasting backgrounds. Ask for a short text block when the image must contain words; keep legal copy, prices and long instructions in an editable layer or document. Generate a small first image, inspect it, then continue the series with the accepted direction.

When a result is inconsistent, keep the prompt and references unchanged, switch from automatic to a fixed model and make one new attempt. Try the general model for a simple subject, sunburst for identity or continuity, and flare for text-heavy layout. If the current catalog does not expose a requested model or quality, use an available option shown in the selector. Treat the failed attempt and the new attempt as separate billable work according to the current pricing page; do not assume a retry is free.

Create [one image / a series of N images] for [purpose].
Model preference: [automatic / fixed model ID].
Canvas: [aspect ratio], [PNG/JPEG/WebP], [opaque/transparent] background.
References: [identity reference] controls [details]; [style reference] supplies [lighting or palette].
Composition: [subject, camera, hierarchy and safe margins].
Text: include only “[short exact wording]”; keep it large and legible.
First make a small draft. Tell me which subject, text and edge details I should inspect before continuing.

A clear checkpoint makes it easier to compare automatic routing with a fixed model and to continue a series without changing its visual contract.