Archive Sora outputs before 24 Sept

Download every Sora file, prompt and job id before the API dies on 24 September 2026. The catalog will not keep them for you. Sora's app and web already shut on 26 April 2026. What still answers is the API: sora-2, sora-2-pro, and snapshots. That surface has a published end date. This page is a copy checklist, not a model-routing guide. Routing lives on the Seedance, Veo and Kling migration posts. Here you make sure the old masters still exist as files you hold.

What dies on 24 September

Three facts, and only these:

Sora web and app were discontinued 26 April 2026.
The Sora API — sora-2, sora-2-pro, all snapshots — shuts 24 September 2026.
Versely's catalog still lists Sora 2 text-to-video and Sora 2 Pro Storyboard as of this writing. After the API dies you cannot re-run those rows against OpenAI. Anything that exists only as a remote job is gone.

Do not speculate about why OpenAI ended the consumer app or the API. The dates are enough to act.

What "dies" means in practice:

You cannot submit a new sora-2 or sora-2-pro job after the cutoff.
You cannot retry an old job id to get a fresh sample from the same seed.
You cannot pull a result URL that has already expired if you never stored the bytes.
Snapshots of those models go with them. A pinned snapshot is not a loophole.

What does not automatically die:

MP4s and stills already on disk, in a bucket, or in a Versely library download you actually completed.
Prompts in your own repo, Notion, or shot list.
Posts already published on TikTok, Reels, Shorts. The platform file is a copy. It is not an archive of the prompt.

Treat 24 September as a hard stop on generation and retrieval, not as a delete of every Sora-looking clip on the internet. The risk is the work you assumed you could "just run again."

Files, prompts, job ids

Copy four objects per job, in this order. If you only copy the MP4, you can post it. You cannot rebuild it.

1. The media. The video file, and the audio if it was separate. For Storyboard, every scene file, not just the concat. If you only ever watched a preview in a browser, that preview is not the archive. Download the result. Name it with date, slug, duration, and a short shot id: 2026-06-14_sora-2-t2v_08s_pack-hero-03.mp4.

2. The prompt. Positive prompt, negative prompt, aspect, duration, any style flags. Storyboard: the ordered list of scene prompts, not a summary. If the prompt lived only in OpenAI's UI and that UI is already gone (app/web, 26 April), you may already have a hole. Search Slack, email, and agent logs now.

3. The job id and request metadata. OpenAI request id, Versely request id if the generate ran through Versely, timestamp, model id (sora-2 vs sora-2-pro vs a snapshot name), seed if you have one. You will not resubmit these after the 24th. You will use them to prove which file is which when legal or a client asks "is this the April take or the June take."

4. The inputs. Reference stills you uploaded, first frames, storyboard boards, driving images. Sora 2 Pro Storyboard is listed as image-to-video. If the still is not in your brand kit, it may exist only on OpenAI's side. Pull it.

Checksum the media (SHA-256 is enough) and store the hash next to the prompt. If you later transcode for a platform, you can still point at the master.

Store this as a folder per campaign, not as a pile of download (17).mp4. A spreadsheet with columns shot_id, filename, sha256, prompt_path, job_id, duration, model, downloaded_at is the index. Astra can fill that sheet if you want computer use for the paperwork. It cannot fetch an expired URL.

What not to archive, and what not to wait on

Do not archive every failed take unless you have a reason. Failures are useful when they document a defect you still need to avoid. They are not a second library. Keep the last good and the last instructive bad. Delete the rest after you are sure the good file is on disk.

Do not archive platform encodes as if they were masters. A TikTok download is compressed, captioned, and sometimes cropped. If that is the only copy, say so in the index (source: platform-export). Prefer the generate output.

Do not wait for a "Sora 3" or a Storyboard clone. This checklist does not depend on a replacement existing. Replacements are other models with other clocks. You still need the old file.

Do not assume Versely will snapshot OpenAI's blob store for you on the 24th. If the bytes are not in your library download, treat them as at risk. Open the library, download, confirm the file plays offline.

Do not spend the week regenerating on Sora "for luck." The API still works until it does not; each generate is credits you could spend on the model you will actually use in October. Archive first. Regenerate later on Seedance, Veo, or Kling only for jobs that must ship as new pixels.

Do not mix this folder with new generates. New Seedance 2.0 takes belong in a different tree. The Sora tree is a frozen archive. Once 24 September passes, it should only grow if you find a disk you forgot.

A one-week checklist

If you are reading this with less than a week left, do it in this order. Skip polish.

Day 1 — Inventory. List every Sora job from production calendars, invoices, and the Versely library filter for Sora 2 text-to-video and Sora 2 Pro Storyboard. Count files you cannot find. That number is the work.

Day 2 — Download. Pull every result that still has a live URL. Play each file. A 0-byte download is not a download. Log failures.

Day 3 — Prompts and ids. Dump logs, agent transcripts, and any CSV of request ids. Match them to filenames. Gaps get a row that says prompt: missing.

Day 4 — Inputs. Brand stills, storyboard boards, first frames. If a still only exists inside a Sora take, export a frame and mark it extracted-from-sora, not brand-master.

Day 5 — Index and checksum. Fill the spreadsheet. Hash the masters. Put a copy of the folder on a second disk or a bucket you control.

Day 6 — Client and legal. If a client paid for a Sora batch, send them their files or confirm they already have them. Do not promise a re-run after the 24th.

Day 7 — Stop generating on Sora. Point new work at the migration rows: Sora 2 to Seedance 2.0, Sora 2 to Veo 3.1, Sora 2 to Kling 3.0. Keep this archive read-only.

If you find a job on the 23rd that was never downloaded, that day is for that file, not for a new campaign. After the 24th, missing jobs are missing.

When the archive is closed, new pixels go through a live row in the studio. The Sora folder does not get another generate. It gets a label: sora-api-ended-2026-09-24.

FAQ

Will Versely re-run my old Sora jobs on 25 September?

No. After the OpenAI API shutdown there is nothing to re-run those model ids against. Downloaded files remain. Undownloaded results do not.

I only have the TikTok post. Is that an archive?

It is a platform copy. Keep it, label it as such, and look for the generate master. If there is no master, that post is the best remaining file.

Should I regenerate everything on Seedance before the 24th?

No. Archive first. Regenerate only what must ship as new pixels. A finished Sora master that you hold is still a finished master after the API dies.