sume/auto for images: no model named, no seed, so pin ids for brand
sume/auto picks an image family and never says which, and there is no seed. When Auto is fine, and when to pin an id like GPT Image 2.5.

model: "sume/auto" lets Image Router choose the family, and Sume never discloses which one ran: it is not listed in GET /v1/images/models, and job.model stays sume/auto. The Image API also does not serve seed today. Together that means you cannot reproduce an Auto result or tell your design team which model made it. Use Auto to explore; pin an id such as openai/gpt-image-2.5 for anything brand-facing.
What the docs say
The Image API page lists Auto as a Sume-only value, notes that model echoes what you requested, and says output_compression and seed are in the schema but advertised by no model, so sending one returns 400 unsupported_parameter. It also says Auto routing continues to use Flare for the GPT Image 2.5 family. The legacy Image 1.0 URLs are compatibility aliases for the same Auto pipe.
| Question | sume/auto | Pinned id such as openai/gpt-image-2.5 |
|---|---|---|
| Which family ran? | Not disclosed | The id you sent |
| Listed in GET /v1/images/models? | No | Yes |
| Parameters you can check beforehand | None published for Auto | supported_parameters on the catalog record |
| Seed for repeatable output | Not served | Not served |
| Best for | Exploration | Brand and repeat work |
Why pinning helps without a seed
A pinned id does not make two generations identical, since there is no seed. It does make them comparable: same family, same limits, same price line. With Auto, a style that worked last week may come from a different family next week and you will not see the change.
Pin the id in config, log it with the job metadata, which Sume stores on the job and does not send to the provider, and record the prompt and the reference URLs. That gives you a lineage you can audit.
{
"model": "openai/gpt-image-2.5",
"prompt": "Product on seamless white, soft shadow, centred",
"quality": "high",
"metadata": { "campaign": "autumn-launch", "variant": "v3" }
}When Auto is the right call
- First look at a new idea, where any decent result will do.
- Throwaway visuals such as internal mockups.
- A pipeline step where you check the output yourself every time.
- Never as the only record of what made a hero image.
Moving from Auto to a pinned id
Run your usual prompt on a pinned id with n: 2, compare, and read supported_parameters for ratios and the reference range before you change anything else. A parameter the pinned model does not list is rejected, not ignored. See Jobs and results for storing the job record.
Sources
Related posts
More in Developers
- Sume bulk items ignore on_active_run skip: allow is forced
Setting on_active_run skip or reject on a Sume bulk item does not stall the window, because the bulk controller runs every item with allow. What that changes.
- Count Sume bulk webhooks to know the queue is done, and the trap
A Sume bulk queue has no webhook, so some teams count per-item deliveries. Canceled and skipped runs never deliver, so a pure counter can hang. Use a hybrid.
- Run sume doctor --agent --json as a CI preflight and keep the output
sume doctor --agent --json checks local CLI readiness without calling the API. Run it before job steps, fail on bad JSON or a nonzero exit, and keep the file.
- Sume Format run status is spelled canceled with one L: fix your switch
A Format run ends as completed, failed, canceled or skipped. If your code matches cancelled with two Ls, a canceled run will look like it never ended.
Written by Sume