sume/auto or a pinned video id after a vendor retires a model?

Pin a catalog id when you must know which model made a clip, and use model: "sume/auto" when you only need a good clip and want Sume to keep choosing. Sume's docs say a replay of an sume/auto request gets the same price and the same route, and that the response reports sume/auto without naming the family.

OpenAI removed the Videos API and every Sora 2 id on 2026-09-24 (its deprecations page, read 2026-10-07), which is the clearest recent case for deciding this before a retirement, not during one.

What each choice guarantees

Both go through the same POST /v1/videos request. The difference is only the model field.

A decision rule that fits on a card

If a client has ever asked which model made a shot, pin. If a clip is throwaway, such as a first-draft hook or a social variant you will test and discard, use Auto. If you do both, keep two config values and write the submitted model string on every job record.

• Pinned ids give the reproducibility an approval trail needs.
• Auto lets Sume change the family without a code change on your side.
• Neither removes the need to read the response status and handle failed.

A Python submit that takes either

This sketch reads the model from the environment, so switching between sume/auto and a pinned id is a config change. It follows the documented submit and poll flow.

import os, time, requests

BASE = "https://api.sume.com"
H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}

def main():
    body = {
        "model": os.environ.get("VIDEO_MODEL", "sume/auto"),
        "prompt": "A slow push-in on a ceramic mug on a wooden desk",
        "duration": 5,
    }
    job = requests.post(f"{BASE}/v1/videos", headers=H, json=body, timeout=30).json()
    while True:
        time.sleep(30)
        s = requests.get(job["polling_url"], headers=H, timeout=30).json()
        if s["status"] in ("completed", "failed", "cancelled"):
            print(s["status"], s.get("model"))
            break

main()

What to log

Log the requested model, the model in the poll response, the job id and usage.cost. With a pinned id the two model values match. With Auto the response says sume/auto, so keep your own record of why that choice was made.

Where Auto helps and where it hides things

Auto removes a decision, and that is both its use and its cost. You get a working clip without choosing a family, and a later change in what Sume routes to does not need your code to change. You lose the ability to say which family produced a given clip, since the docs state Sume never discloses it and tell callers not to infer it from the output.

That makes Auto a good fit for volume work where the clip is disposable and a poor fit for anything that is billed back to a client by model, or compared across a test run. If you are running a bake-off, pin every row; otherwise you are comparing Sume's routing, not the models.

• Bake-offs and audits: pin.
• Disposable variants and drafts: Auto is fine.
• Look-matched campaigns: pin, and record the id with every job.

Cost and limits with Auto

Auto's create controls default to 720p and 8 seconds, with clips of 3 to 10 seconds in 16:9 or 9:16. If your brief needs 15 seconds or a square frame, Auto is outside its envelope and you must pin a row that supports it. The pinned rows list their limits in the catalog, and the price of each is its list rate times 1.25 per second.