How long AI video vendors keep your file: Veo 2 days, Higgsfield 7+

Vendors differ by orders of magnitude on how long a generated video stays downloadable: two days for Veo on the Gemini API, a minimum of seven days for files on Higgsfield, and 24 hours for Sora Batch API outputs. The safe rule is the same everywhere: download or copy the file when the job completes.

What each vendor states

Retention is a line in each vendor's docs, usually far from the pricing table. The values below are what each page stated on 2026-10-03.

Why the windows matter

Two of those three are windows you can miss in a normal week. A reviewer who is out on Monday can lose a Friday render on Veo. A batch queue that finishes overnight is a risk if nobody downloads before the window closes.

What Sume does with outputs

Sume mirrors generated outputs into Sume-owned media URLs and tells integrations to store the Sume URL rather than raw provider URLs. See the media inputs docs. The pages read here do not state a retention period for Sume artifacts, so copy the file to your own storage anyway if you have a retention rule of your own.

Copy on completion

A completion handler that copies the file the moment the job is done removes the window problem, whichever vendor serves the model. The poll response for a Sume video job carries unsigned_urls, and the content endpoint accepts your API key.

import os
import time
import requests

headers = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}


def wait_and_save(polling_url: str, path: str, interval: int = 30) -> None:
    while True:
        job = requests.get(polling_url, headers=headers, timeout=30).json()
        if job["status"] == "completed":
            video = requests.get(job["unsigned_urls"][0], headers=headers, timeout=120)
            video.raise_for_status()
            with open(path, "wb") as f:
                f.write(video.content)
            return
        if job["status"] in ("failed", "cancelled"):
            raise RuntimeError(job.get("error", job["status"]))
        time.sleep(interval)

A retention policy for your own pipeline

Whatever the vendors do, decide what you keep. A short policy avoids both lost renders and unbounded storage. Keep the approved final and its job record for as long as the ad runs, plus a margin. Keep rejected takes for a few days to settle disputes about why one was chosen, then delete them.

The Sora Batch figure is historical, since the Sora API shut down on Sep 24, 2026. A pipeline that copies on completion does not depend on any of the windows above.

• Finals: keep with the job id, model id, and prompt.
• Rejected takes: keep briefly, then delete.
• Copy on completion, never on review day.
• Test a restore from your own storage once a quarter.

Add a webhook

Pair the copy with a webhook so nobody has to poll. The video docs describe callback_url for that, and the stored post on Veo's two-day storage covers the Google side in more detail.