Sora Is Gone. What a Production Should Have Kept, and What to Ask For Next

What happened to Sora?

OpenAI closed Sora in two steps. The Sora web and app experiences were discontinued on 26 April 2026, and the Videos API and all Sora 2 models were removed from OpenAI's API on 24 September 2026. OpenAI said user data would be permanently deleted after the export deadlines, and named no replacement model.

The dates come from OpenAI's own API deprecations page, which says developers were notified on 24 March 2026 that the Videos API and the `sora-2` and `sora-2-pro` aliases and snapshots would be removed on 24 September. The same day, NBC News reported OpenAI's post on X — "We're saying goodbye to Sora" — and described the move as a shift of computing resources towards text and code ahead of an expected IPO, citing Wall Street Journal reporting.

The collateral was not small. On 11 December 2025 Disney announced a $1 billion equity investment in OpenAI and a three-year licence for more than 200 Disney, Marvel, Pixar and Star Wars characters on Sora. NBC News reported that the investment and partnership would not go ahead, and that Disney said it would respect OpenAI's decision.

We covered where generative video fits in a pipeline in August, when the sunset was announced. This piece is about the other half: what a production that actually used a hosted model should do now, and what it should ask of the next vendor.

What happens to data after a hosted model is switched off?

On the evidence of Sora, it goes away. The Decoder, summarising OpenAI's Help Center guidance in March, reported that users were told to download videos and images from the Sora library before the cutoff, that user data would be permanently deleted once the deadlines passed, and that OpenAI had not yet decided whether to offer a final export window. If a producer, an editor or a freelance previs artist generated work on a personal account and did not export it, assume it no longer exists.

The more uncomfortable point is that deletion is the normal case, not the shutdown case. Google's Gemini API documentation says generated Veo videos "are stored on the server for 2 days, after which they are removed." Runway's API documentation says its output URLs "will expire within 24-48 hours" and that it expects you to "download the data at this endpoint and save it to your own storage." A hosted video model is not an archive, even on a good day. Sora just made the point with a calendar.

A vendor-risk checklist for producers and post supervisors

The post-production supervisor is usually the person who ends up owning this, because they own the deliverables list. The checklist works just as well for a producer signing off a pitch or a previs budget.

1. Inventory. List every shot, board, pitch clip and temp element that came out of a generative tool, who generated it, and on which account. Personal accounts are the most common hole.
2. Pull everything down. Download at the highest quality offered, straight into the production's own storage, not a crew member's laptop.
3. Record how it was made. Prompt, references, model identifier, settings, seed if exposed, date. Without these you cannot reproduce, explain or defend the shot.
4. Mark where it lives in the cut. A temp element that reached the edit needs to be flagged in the EDL or the VFX shot list, so nobody forgets it was never meant to ship.
5. Classify each use. Disposable (previs, boards, lookdev) or delivered (anything in a master). Only the second needs the full rights paper trail.
6. Price a fallback. For every generated element that is still in the cut, carry a line for replacing it by another route — a different model, a VFX vendor or a reshoot.
7. Diarise deprecation notices. Somebody on the production should be subscribed to each vendor's deprecations or changelog page. OpenAI gave six months; the next vendor may give less.

What to archive, and where

That last row is worth a sentence. Google says Veo output is watermarked with SynthID and can be checked through its verification platform. The C2PA standard behind Content Credentials describes itself as "a nutrition label for digital content." Both only help if your pipeline keeps the original file intact somewhere, because a re-encode can drop embedded metadata.

Contract clauses to ask for

None of this is legal advice. It is the list of questions to take to your lawyer before a generated element goes anywhere near a deliverable, and it applies both to the tool's terms and to any contract with a vendor or freelancer who uses one on your behalf.

• Output ownership. Who owns what the model produces, and does the vendor keep any licence to it?
• Training use. Are your prompts, uploads and outputs used to train the vendor's models, and can you opt out in writing?
• Retention and export. How long are outputs kept, and in what formats can you export them?
• Deprecation notice. A minimum notice period before a model or endpoint is removed, and a guaranteed export window after.
• Indemnity. Whether the vendor stands behind its output if a third party claims infringement, and what caps apply.
• Provenance. Whether outputs carry a watermark or Content Credentials, and a promise not to be asked to strip them.
• Disclosure down the chain. Freelancers and vendors must tell you when a deliverable contains generated material and hand over the records in the table above.

For the credit and copyright side of this, see AI screenwriting credit and copyright; the same questions about human authorship apply to images.

How to evaluate a replacement

Sora's API customers are now choosing again. Google's Gemini API currently lists Veo 3.1 models, and Runway's API documentation lists models including Gen 4.5 for image-to-video and Aleph 2.0 for video-to-video. We are not ranking them; quality moves too fast to be worth printing. What stays useful is the order of questions:

1. Can we get our work out? Retention period, export formats, bulk download.
2. What do the terms say about rights and training? Read the version in force today, and save it.
3. Is provenance built in? Watermark, Content Credentials, or neither.
4. What is the deprecation record? Does the vendor publish a deprecations page, and how much notice has it given before?
5. Only then: does it look right? Run the same archived prompts and references through it. If you kept the records above, this test takes an afternoon.

What this does not tell you

We do not know how many productions used Sora in a delivered master, and we found no reliable figures on what it cost OpenAI to run; we have left both out rather than guess. The export-window detail comes from press reports summarising OpenAI's Help Center, which we could not load directly. Retention periods and model line-ups quoted here are from the vendors' own documentation as of late September 2026 and can change without notice. And whether a given generated shot can be copyrighted, licensed or delivered to a particular distributor depends on the jurisdiction and the contract. Ask your lawyer.

The wider picture of AI in film production has not changed because one model closed. What has changed is that "the vendor will keep it" is no longer an assumption anyone can make with a straight face. Keep the records in your own post-production system, and the next shutdown is an inconvenience rather than a hole in the cut.