Most one AI video request can cost on Sume, by model

The most one Sume video request can cost is set by two limits: the longest duration the model accepts and the dearest resolution it offers. Seedance 2.5 sits at the top of the router catalog at $42.65 for 30 seconds at 1080p; Wan 3.0 stops at $7.50 for the same length. Knowing the ceiling per request is what lets you choose a spend cap that is useful, rather than one that is either too tight to deliver or too loose to protect you.

Limits are from the Video Router docs and Generate avatar video; prices are from the API pricing rate card, read on 2026-10-03.

What is the ceiling for each router model?

The table multiplies the longest, sharpest request each entry accepts by its Sume rate. Video-to-video entries (Genjutsu and H3 Max Recast) bill by the length of the source clip you send, so their ceiling is reached by uploading a 30-second source.

And for Avatar Video?

Avatar Video accepts scripts whose estimated duration is 4 to 60 seconds, billed per second by tier, with a slightly higher rate when a product image is attached. Quality defaults to plus when you omit it, so the default ceiling is $14.70, not the standard-tier $11.04.

Where do the spend caps bite?

A scheduled run carries a default generation cap of $1.00, so on a schedule one 30-second Wan 3.0 clip at 1080p does not fit under the default, and neither does a 15-second 1080p Seedance 2.0 clip. A Format run inherits its Format's cap, or $400 when the Format never named one, and a request can name its own up to $500 (Create a run). Agent Completions require generation_spend_cap_usd and have no default (Agent Completions).

Two cap sizes cover most work. For one Wan 3.0 or H3 clip, a cap around the table value plus a retake is plenty. For Seedance 2.5 at 1080p, a single 30-second request is $42.65, so a $40 cap cannot finish it; the run fails rather than overspending.

Does the wallet need the whole amount up front?

Yes. Sume reserves the estimated amount when it accepts a paid job, and a balance too small to reserve it fails with 402 insufficient_credits before provider work starts (Generation admission). Check GET /v1/balance before you submit the expensive end of the table, and see what a failed job charges for how the reserve is released.