Which model did my AI video use? sume/auto does not say

You cannot find out which model made a sume/auto video, because Sume does not disclose it. The docs say responses echo sume/auto and that the family that ran is never disclosed. The only way to know the model is to send a catalog id yourself, such as seedance-2.5, so the request names the model up front.

What does the job tell me?

The poll response reports "model": "sume/auto". Nothing in it names a family. The MCP instructions say the same for agents: sume/auto echoes back verbatim and the family is never disclosed. The Video generation docs add that you should not build on any observable trait of the output to infer the family, so do not guess from the look of the clip.

Why does that matter for my workflow?

Three cases come up. A reviewer asks which model produced a shot. A style has to match an earlier clip. A limit depends on the model, such as length. Auto cannot answer the first, cannot promise the second, and only documents one length range for all of them.

How do I pin a model instead?

Send the catalog id as model on POST /v1/videos. The docs' own example uses seedance-2, and the response echoes that id. Read the id list from the video models endpoint, or from video-router_models over MCP, and store the id you used next to the job id in your own records.

For an agent, name the family in the request: over MCP, generate_video routes to sume/auto only when payload.model is omitted.

What should I record for my own audit trail?

Record what you sent, not what you suspect. Keep the job id, the model value from your request, the full body, and the Idempotency-Key. For an Auto job the honest value is sume/auto. Do not write a guessed family into a report or a client deliverable, because nothing in the response supports it, and the docs tell you not to infer it from the output.

If a client contract requires naming the model behind each clip, treat Auto as out of scope for that work and pin an id from the start.

Does a replay pick the same model?

The docs say resolution is a pure function of the normalized request and the catalog version, so an idempotent replay prices and routes identically. That is a promise about repeats of one request, not a disclosure. It does not tell you which family the first call used.

For the wider comparison, read sume/auto or a pinned video model and what Sume's router does.