Most expensive single AI video request on Sume: max cost per model

If you let an agent or a script submit video jobs on its own, the number that matters is the worst case for one request. Sume reserves the billable amount on submit, so a single job can hold up to the figures below before it settles. This post computes that ceiling for every row with a published per-second rate: the longest duration the row accepts at its most expensive resolution.

The source-driven rows, Recast and Genjutsu, are the outliers. Their rates are several times higher per second, and the length comes from the source clip, so the cap is on how long a source you pass.

What is the ceiling per row?

Rates are the provider list times 1.25, rounded up to the cent per job. The Seedance rows bill per video token and Grok Imagine has no rate in this table, so neither is shown.

How do I cap spend in practice?

Check the balance rules before a batch: funds are reserved at submit, so a workspace with less than the reserve gets a 402 insufficient_credits instead of a job. For automated callers, pin the model and resolution in your code rather than letting a field default decide, and clamp duration to a number you chose.

The catalog also tells you the limits. A value outside supported_durations or supported_resolutions is a 400, so a bad request cannot silently turn into a long, expensive one.

Is the reserve the same as the final bill?

The reserve is an estimate from the request body; the final amount is usage.cost on the finished job. For text and image requests with a duration you set, the two match. For Recast and Genjutsu the duration is the inspected source length, so inspect first and send that number, as described in the Recast rounding post.