Flow Agent picks the image model for you: Sume's sume/auto does too

Flow has two ways to make an image: the Agent, which routes your request to the best image model, or the standard prompt box where you choose the model yourself, with Nano Banana Pro as the default (read 2026-10-02). Sume has the same split. sume/auto lets Sume pick a family and never tells you which; a catalog id such as google/nano-banana-pro pins it.

Flow's two paths

Flow's help page says you can use the Google Flow Agent, which automatically routes requests to the best image model, or the standard prompt box with manual model selection (Flow Help, read 2026-10-02). The page does not say which models the Agent can choose from.

Sume's two paths

Send model: "sume/auto" to POST /v1/images and Sume picks the family. It is not listed in GET /v1/images/models, the job model stays sume/auto, and the family that ran is not disclosed (Image API). Send a catalog id instead, for example google/nano-banana-pro, and that is the model that runs.

When automatic routing is a poor fit

A set of images that must look like one family, such as a product line or a character sheet, needs one model. Auto may route two requests differently, and Sume tells you nothing you can build on about which one ran. For a series, pin the id and keep your reference images identical across calls; input_references takes up to 10 on that model.

When automatic is fine

For a single throwaway visual or a first draft where any decent result will do, Auto saves the model research. It also keeps working when the catalog changes, since you named no id that can be retired.

Check the response: usage.cost is the amount billed, so a pinned versus Auto comparison on your own prompts costs little and shows which you prefer.