Image Router catalog quote: 23 cents estimated, 25 applied, 777 cap

The short answer

The public catalog row for the Image Router (read 2026-10-09) gives an estimated price of 23 cents, a minimum of 1 cent and a maximum of 736 cents. It also lists applied figures of 25, 2 and 777 cents, where the difference is an agent fee set at 550 basis points (5.5%). So a single Image Router call can be reserved anywhere from 1 cent to $7.36, and up to $7.77 with the fee.

The fields

Open https://api.sume.com/v1/catalog without a key and find the entry with id image-router. Its model_pricing block has the fields in the table. The reservation text says Sume stores the estimated billable amount in USD micros at submit time, rounds up to cents, captures on completion, and refunds failed or pre-generation cancelled jobs before capture.

Checking the arithmetic

The fee is 550 bps, which is 5.5%. On the 23-cent estimate that is 1.265 cents, which the catalog shows as 2 after rounding up to a whole cent, giving 25. On the 736-cent maximum, 5.5% is 40.48 cents, and 736 + 40.48 = 776.48 rounds up to 777. On the 1-cent minimum, 1.055 rounds up to 2. The catalog row does not say which kinds of call pay the fee, so check a real usage row before you build a budget on the applied figures.

Reading it in code

The catalog is public and needs no key. A few lines of Python can fetch it, find the entry whose id is image-router, and print model_pricing[0]. Run that check on the day you plan a large job, since the figures are a snapshot and the catalog can change. If the maximum rises, your worst-case hold per call rises with it. For a batch of 50 calls the worst-case hold at the maximum would be 50 x $7.36 = $368, which is why naming a model and a quality tier matters.

When to use the router

The router makes sense when you do not want to pick a model. If you already know the model and the quality tier, call the model directly: the price is then a single known figure rather than a range from 1 to 736 cents. The router's range is wide because it covers every model and size it may choose from.

Treat the maximum as the amount a single call might hold, not as what it will cost.

The 23-cent estimate is a mid-point for the whole router, not the price of any model. Real prices are per model: for example $0.025 for grok-image and $0.2225 for gpt-image-2.5 at high 4K. The reservation is a hold, and the docs say you pay cost_usd times n for a completed generation.

Two habits keep the hold small. Name a model instead of sume/auto, and set quality and the size yourself. The Image API docs say auto quality reserves max, which is the top of the range.