Let QA override the image model per request, with an allowlist

Add an optional X-Image-Model header to your own internal render endpoint, and honor it only when the value is on a short allowlist of Sume ids. QA can then test openai/gpt-image-2.5-sunburst against real traffic shapes while the production default still points at the old id. When the test passes, you change the default, and the header stays as a debugging tool.

The allowlist matters because the id goes straight into the model field of POST /v1/images. An unlisted id returns 404 model_not_found, and an unchecked header would let any caller spend your budget on the most expensive id.

Pick the model

The function below is the whole policy. Put it in the handler of your own service, ahead of the call to Sume.

ALLOWED = {
    "openai/gpt-image-2.5-sunburst",
    "openai/gpt-image-2.5",
    "google/nano-banana-2",
}
DEFAULT = "openai/gpt-image-2.5-sunburst"

def choose_model(headers):
    asked = headers.get("X-Image-Model")
    if asked is None:
        return DEFAULT
    if asked not in ALLOWED:
        raise ValueError(f"model {asked!r} is not on the allowlist")
    return asked

if __name__ == "__main__":
    print(choose_model({}))
    print(choose_model({"X-Image-Model": "google/nano-banana-2"}))
    try:
        choose_model({"X-Image-Model": "gpt-image-1"})
    except ValueError as err:
        print(err)

Keep the allowlist small

Return the chosen id in the response so QA can see what ran. Do not put sume/auto on the list if the point is a controlled comparison: Sume never discloses the model behind it. The config-file side of this change is in the model-map post, and the traffic-split variant is in the canary post.