Decart lucy-latest vs a pinned Lucy model; Sume catalog ids

Decart lets you call a Lucy model by a pinned id such as lucy-2.1 or by a latest alias such as lucy-latest, and its price list shows why the choice matters: Lucy 2.5 and 2.1 are $0.04 per second, while the legacy Lucy Clip is $0.15 per second and capped at 5 seconds. On Sume the equivalent habit is to pin a catalog model id and read its limits from the catalog, rather than assume them.

Decart's details come from its Video Editing page and Python Queue page, read 2026-10-02.

What does Decart's price list say?

The editing page lists three models. Lucy 2.5 is recommended, unlimited duration, fastest, and $0.04 per second. Lucy 2.1 is the previous generation, with the same duration and price. Lucy Clip is legacy, 5 seconds at most, standard speed, and $0.15 per second. So the legacy model costs 3.75 times as much per second and limits you to a few seconds. The Queue page lists lucy-2.1, lucy-vton-3.5, and lucy-restyle-2 as supported models and names the aliases lucy-latest, lucy-vton-latest, and lucy-restyle-latest.

Why is an alias a risk?

Decart's editing page says lucy-latest points to Lucy 2.5 today, and it recommends Lucy 2.5 (lucy-2.5 or the lucy-latest alias) for new projects. The pages do not say how often the alias moves. If it can point at a different model next quarter, your output style and your bill can change without a code change. For a production job, pin the id, and for a prototype use the alias only if you accept the drift.

How does Sume name models?

Sume's Video Router is an explicit catalog: you pick a model id from GET /v1/video-router/models, and the docs advise reading each model's capabilities instead of assuming one envelope, because limits are per model. Sume also offers Video Router Auto through model: "sume/auto", where Sume picks the model; the docs recommend it for new integrations. Auto defaults to 720p and 8 seconds. The tradeoff is the same one: Auto is convenient and a pinned id is predictable.

Pin when you have tested a model on your footage, and record the id with the job.

curl -s https://api.sume.com/v1/video-router/models/gemini-omni-flash-1.1 \
  -H "Authorization: Bearer $SUME_API_KEY" | jq .capabilities

How do I keep the bill predictable?

Sume bills list times 1.25 on every Video Router model, per the docs, and the rate depends on the model and resolution. Store the model id, the resolution, and the job id with each result. If a month's spend moves, you can tell whether the cause was a different model or a different volume. The same discipline works on Decart: pin lucy-2.1, not lucy-latest, wherever you need the number to hold still.

What should a config file record?

Keep four fields next to every generated clip: the model id you requested, the resolution, the job id, and the date you last read the vendor's price page. When a vendor changes a model or a price, those four fields let you explain a different result or a different invoice without guessing. For Decart that means the pinned Lucy id and the per-second rate; for Sume it means the catalog id, the resolution, and the job id.

Re-read both vendors' pages before a big batch. The figures here are dated 2026-10-02, and neither Decart nor Sume's docs promise them for next month.