Image Router reserves 23 cents, catalog max $7.36: wallet needed

The Sume catalog row for Image Router estimates 23 cents per call, with a minimum of 1 cent and a maximum of 736 cents ($7.36), read 2026-10-09. The estimate is what a reservation uses when the exact model is not known, so a wallet needs to cover at least the estimate and, for the largest requests, up to the maximum.

What the catalog says

The catalog entry for sume/image-router publishes a price estimate in cents. The text beside it says that Image Router pass-through models reserve from provider list pricing times a 1.25 Sume margin, and that the per-model rates come from the models endpoint. It also says that a reservation is stored at submit time, rounded up to cents, captured on completion, and refunded if the job fails or is cancelled before capture.

Why sume/auto does not have one price

With model: "sume/auto", Sume chooses the family and never discloses which one ran, and job.model stays sume/auto. Because the model is not fixed, the Image API cannot show you one per-image row ahead of time. The catalog therefore gives a range, and the estimate of 23 cents sits between the cheapest rows (a few cents) and the expensive ones (the 4K rows run to $0.375 for Nano Banana Pro).

For predictable spend, pick a model id and read its price from GET /v1/images/models. For example, Nano Banana 2.1 at 1K is a fixed $0.10 row, whereas Auto's final amount depends on what it selects.

Sizing a wallet

Reservations hold money until a job finishes. If you submit 20 Auto jobs in a burst, the held amount at the 23-cent estimate is 20 x $0.23 = $4.60. If you size for the maximum, 20 x $7.36 = $147.20, which is a ceiling, not a typical bill.

A reasonable rule: keep the wallet above the estimate times your in-flight jobs for normal use, and pin a model when you need a tight ceiling. Read errors and credits for what an insufficient balance returns.

Pinning a model

If you need to keep the reserve small, name the model. A request pinned to GPT Image 2.5 at low quality reserves from a $0.02475 row, a tenth of the Auto estimate, while a pinned Nano Banana 2.1 request at 1K reserves $0.10. Auto remains useful when you want Sume to choose and you accept a wider range of cost.