sume/auto or a fixed video model id: decision rules for client work

Use sume/auto when you want a good clip and do not care which family makes it, as long as the clip is 3 to 10 seconds. Pin a model id when the job needs something auto cannot promise. Per the docs, the family behind sume/auto stays opaque, responses show sume/auto, and when the router cannot honor a request it fails closed with 400 unsupported_capability instead of quietly picking something worse.

Rules that pin an id

Pin when any of these is true:

• The clip is longer than 10 seconds: Seedance 2.5 and Wan 3.0 reach 30 s, Kling 3 and the H3 rows reach 15 s.
• You need 4:3 or 3:4: sume/auto clips are 16:9 or 9:16 only.
• The client signed off on a family and a repeat run must match.
• You need a source-clip row such as Recast or Genjutsu, which auto never selects.
• You need a silent clip: auto clips always carry native audio.

Rules that keep auto

Keep auto for drafts, internal tests and short social clips, and when a launch changes the best row: the default can improve without a code change on your side. Log the model that actually served the job, so a client can see what they got.

Pricing follows the row: Sume bills list rates multiplied by 1.25, per second for most rows and per 1000 video tokens for Seedance.