Detect new Sume video models daily: a Python catalog diff

Short answer

Call GET /v1/videos/models on a schedule, save the result, and diff ids against the previous snapshot. The response is a data array, and each model has an id, supported_durations, supported_resolutions and supported_input_references per the video docs. A new id means a new model; a missing id means a model went away or its provider is not configured. The script below prints both.

The script

It stores the snapshot as video-models.json next to the script, so the first run reports every id as new. Run it daily from cron or a scheduled job.

import asyncio, json, os, pathlib, urllib.request

SNAP = pathlib.Path("video-models.json")

def fetch():
    req = urllib.request.Request(
        "https://api.sume.com/v1/videos/models",
        headers={"Authorization": "Bearer " + os.environ["SUME_API_KEY"]},
    )
    with urllib.request.urlopen(req) as r:
        return {m["id"]: m for m in json.load(r)["data"]}

async def main():
    now = await asyncio.to_thread(fetch)
    old = json.loads(SNAP.read_text()) if SNAP.exists() else {}
    for mid in sorted(set(now) - set(old)):
        print("NEW", mid, now[mid].get("supported_durations"))
    for mid in sorted(set(old) - set(now)):
        print("GONE", mid)
    SNAP.write_text(json.dumps(now))

asyncio.run(main())

Why a diff beats reading blog posts

In launch weeks, announcements run ahead of access. Kling 4.0 Flash is a closed beta and full Kling 4.0 is expected in October, so a post saying it exists is not a statement that you can call it. The catalog is the only list that says what your key can submit to today.

What to watch

Steps

• Put the script on a daily schedule and send the output to a channel your team reads.
• On NEW, run one 480p draft before any production traffic.
• Keep a pinned id for production and use the diff only for discovery.
• Pair it with a dated price check: new models carry their own rate, which Sume bills at provider list times 1.25.

Making the alert useful

Send the output to a channel with the id, duration range and resolutions, not only the word NEW. Add the date and your key's workspace name so the message is self-explanatory a week later. If you run several environments, snapshot each separately: the catalog can differ when a provider is configured in one and not in another, which is exactly how higgsfield-genjutsu appears and disappears.

Caveats

The diff only compares ids, so compare the fields you care about if limits matter. Do not auto-switch production to a new id because it appeared; that is a decision, not an event.