Export the Sora library before delete

Update, 8 October 2026: the Sora 2 API shut down on 24 September 2026 as scheduled. Check OpenAI's Help Centre for whether the library export below is still open before you rely on it.

Export the Sora library at sora.chatgpt.com/sunset before OpenAI deletes Sora-associated data. Click Export. Wait for the email. Keep the masters you might restyle. Credits remaining on ChatGPT/Sora go to Codex, not to more video.

OpenAI’s Help Centre (updated September 2026) is the source, not a screenshot of a feed. The Sora web and app experiences discontinued on 26 April 2026. The Sora API discontinues on 24 September 2026. After discontinuation, and after any final export window passes, OpenAI permanently deletes data associated with Sora use. The Help Centre recommends exporting as soon as possible. If a final export window is offered, it says it will notify by email. That sentence is not a reason to wait.

This page is the export path. It is not the last-day API checklist and not the migration plan. Files first. Routing second.

The Help Centre export path

The official steps are short on purpose.

1. Go to sora.chatgpt.com/sunset.
2. Click Export.
3. Wait for the email that says the archive is ready.
4. Download it to a disk you control, not to a vendor preview cache.

That is the whole door. There is no second secret URL in the Help Centre article. There is no “export only the liked posts” control described there. Treat Export as the full Sora-associated package OpenAI is willing to give you, then sort it yourself.

Do this even if you already downloaded clips in April. The April 26 consumer shutdown and the September 24 API shutdown are different clocks. Files generated through the API after the app died can still sit only in OpenAI’s Sora-associated store. A Desktop folder from March is not the September library.

Do this even if you generated on Versely. Sora 2 Text to Video (32 credits minimum, T2V, native audio, 720p, 4 / 8 / 12 / 16 / 20s) and Sora 2 Pro Storyboard (240 credits minimum, I2V, native audio, 10 / 15 / 25s) were catalog rows that called the same family until the API closed, and Versely has since removed them. A Versely download of a take you ran there is yours. It does not export the rest of the Sora library you built in ChatGPT, the API console, or another host.

If the sunset page errors, retry from a logged-in session that actually owns the generations. Do not invent a third-party scraper. The Help Centre path is the one that matches the deletion notice.

When the email arrives, download the same day. An unread export email is not an archive.

What OpenAI will delete, and what credits do

The Help Centre’s deletion sentence is not a recycle bin. After Sora is discontinued, and after the period of any final export window (if offered) passes, OpenAI permanently deletes data associated with your use of Sora. Prompts, outputs, project furniture, and whatever else it classifies as Sora-associated go with that sentence. Plan as if you will not be able to browse the library in October.

That is why the master file matters more than the prompt. A prompt can be retyped. A 20-second take whose face, pack, and room already survived review is an asset. Regenerating it on another family is a new identity unless you keep the file and restyle from it.

Credits are a separate object. The Help Centre’s refund pointer is the ChatGPT subscription refund article. The credit sentence is specific: purchased ChatGPT/Sora credits can still be used for Codex if you choose. That is not a video generate. It is not a Sora extension. It is not a coupon for Veo. If you still have a ChatGPT/Sora credit balance on the 24th, the remaining useful video action is zero. The remaining useful file action is the export.

Do not delay the export to “use up” video credits on OpenAI’s side. A last generate you forget to download is worse than an unused credit that can only buy Codex. The last-day checklist is stop prompting; this page is keep the files.

Versely prices are credits on Versely rows. OpenAI’s ChatGPT/Sora credit balance is OpenAI’s. Do not convert one into the other in a spreadsheet and call it a rate.

Keep masters you might restyle

Not every Sora file is worth keeping. Keep the ones that are cheaper to continue than to invent.

Dialogue heroes. Takes where the voice and the mouth already matched well enough to ship. You will not get that exact performance back from a new text prompt on another model. Keep the file. Caption it. If you need a new line, that is a lipsync or a new short take on Veo 3.1, not a hope that Flux 3 “remembers” Sora.

SKU plates. Pack shots, labels, and product orbits that legal already signed. The stills that fed Sora 2 Pro Storyboard are as important as the videos. Restyle from the still on Seedance 2.0, Seedance 2.0 Fast R2V, or Veo first-last. A new T2V of “our blue bottle” is how the label lettering becomes a cousin.

Duration outliers. 16s / 20s T2V and 25s storyboard files are longer than Veo’s 8-second ceiling. Keep them. You can cut them. You can extend a different family from a still. You cannot ask Veo 3.1 for a 20-second single pass.

First and last frames. Grab stills from keepers even if OpenAI’s archive is video-only. Those frames become inputs for Veo first-last (from 64 credits, 4 / 6 / 8s, 4K, native audio) and Flux 3 first-last (from 34 credits, 5–20s, native audio).

Discard freely: obvious misses, duplicate seeds, takes you would not show a client. Deletion on OpenAI’s side does not require you to hoard noise. It requires you not to lose the approved object.

Store the keepers next to a one-line brief: duration, aspect, whether audio is native in the file, whether a still exists. That line is what you hand to the AI video generator in October.

After the archive lands

Sort before you generate replacements.

Make four folders: talking, product, drafts, stills. Talking goes toward Veo 3.1 or another native-audio hero when you need a new line. Product goes toward Seedance 2.0 (from 44 credits, 4–15s, 1080p, native audio, T2V+R2V) or Fast R2V (720p, from 55 credits, was 78) when the SKU must hold. Drafts that were only composition tests can be remade silent on Kling 3 Turbo (3–15s, 1080p, audio false, from 22 credits, was 27) and voiced later. Stills feed I2V and R2V rows.

Then caption what you will still ship from the Sora file itself. The model sunset does not sunset burned-in captions. Add captions to video and the AI caption generator take the file you kept. A muted feed does not care that the API died.

If a keeper is photoreal and you publish in the EU, Article 50 deployer labelling still applies at first exposure. The sunset is not a disclosure holiday. Use the Article 50 disclosure template on the file you actually upload, including an exported Sora take.

Do not rebuild the library from prompts “to have a native Versely version” of every clip. Rebuild the jobs you still have to run. The archive is the backup of identity. New generates are for new shots.

When the zip is on your disk, run the surviving rows from Versely.

FAQ

Where is the official Sora export?

sora.chatgpt.com/sunset, then Export, per OpenAI’s Help Centre (updated September 2026). You receive an email when the archive is ready. That is the path to use. Do not wait for a promised extra window if you can export today.

What happens to leftover ChatGPT/Sora credits?

They can still be used for Codex, according to the same Help Centre article. They are not a Sora video balance after the product is gone. Export files. Do not try to convert those credits into Versely credits or into Veo.

Will OpenAI keep my Sora videos after 24 September 2026?

The Help Centre says that after discontinuation and after any final export window, OpenAI permanently deletes Sora-associated data. Plan on deletion. Keep your own masters. Treat anything you did not export as gone.