Log the exact video model id per shot after Runway model removals
When a vendor retires a video model, a logged model id turns the audit into a query. Record it on every shot, and know what Sume echoes.

Log the model id, resolution, duration and job id for every shot you generate. When a vendor removes a model, you then find affected shots with one query instead of re-reading prompts. Sume's poll response echoes model, with one trap: if you send sume/auto, the response says sume/auto and the family that ran is never disclosed, so pin a model id when you need a record.
Vendors such as Runway retire model ids from their APIs from time to time. Sume details here come from the Video Generation docs.
What should one shot row hold?
Keep it boring. Everything below is a field Sume returns or accepts, so nothing needs to be inferred later.
| Column | Source |
|---|---|
job_id | id in the submit response |
model | Echoed in the submit and poll responses |
resolution, duration, aspect_ratio | Your request |
usage_cost | usage.cost in the completed poll response |
idempotency_key | The header you sent |
status | pending, in_progress, completed, failed, cancelled |
What does a removal look like on Sume?
Sume uses bare catalog ids and says an unknown id is a model-not-found case, to be checked against the Video Models API. If a model is retired, a request that names it fails instead of switching silently. That is a reason to log pinned ids: your history shows which shots used the retired model, and your next call can be pointed at a listed replacement after you check GET /v1/videos/models.
A minimal query when a model is retired
With the table above in any SQL store, the audit is one query: select the rows where model equals the retired id and status is completed, then regenerate or flag each shot. Keep the original prompt and the Idempotency-Key in the row, because replaying a key returns the original job; use a new key for the regenerated shot so it is billed as a new job. Failed jobs are described by Sume as releasing the reservation, so they do not need follow-up.
Why not leave it on auto?
sume/auto is convenient for tests, and Sume says resolution is a pure function of the request and the catalog version, so a replay prices and routes the same. It still hides the family, so a log full of sume/auto cannot answer which model made a shot. Use auto for exploration and a pinned id for anything you will ship or have to redo.
Limits: this post does not claim how Sume announces retirements, because the docs I read do not describe a notice process. Watch the changelog and the model list.
Sources
Related posts
More in Use cases
- Meta AI Reels translation: who can use it, and the 1,000-follower rule
Per Meta, Reels translation is free for Facebook creators with 1,000 followers or more and for all public Instagram accounts where Meta AI is available.
- Translate text inside an image: one Nano Banana 2 edit per language
Google says Nano Banana 2 can translate and localize text within an image. Loop one edit per language through Sume POST /v1/images and keep the layout fixed.
- Pin a product card on a live-selling clip with one overlay call
POST /v1/timeline-1.0/compose with operation overlay pins a product still at bottom, center or top of a live-selling clip for $0.02. Layout keys and limits.
- Podcast audiogram API: audio plus a cover still in one render
Make a podcast audiogram video from an audio clip and a cover image with Timeline 1.0: spine, still slot, square output and the cost per minute.
Written by Sume