Re-run an old Sora project
Export the files, rewrite the prompt for the destination model, lock identity with stills. Do not expect the seed to transfer.
You cannot re-run an old Sora project by pasting the prompt into another model and hoping the seed follows. Export the files first. Inventory each shot as a job. Rewrite the prompt for the destination row. Lock identity with stills. Treat the seed as local to the model that produced it.
The Sora app is already gone (26 April 2026). The API goes on 24 September 2026. OpenAI's help article tells you to export at sora.chatgpt.com/sunset. Do that before you "try Veo on the old prompt." A replacement generate will not reconstruct a take you forgot to download, and it will not honour a Sora seed, a Sora cameo token, or a Sora storyboard identity manifold.
Export first, then inventory the job
A "Sora project" in a real folder is usually several different jobs wearing one name. Before you touch a live model, split the folder:
- Finished files you still need. Export. Store them outside ChatGPT. These are masters, not prompts.
- Prompts that produced a look you liked. Keep the text. Do not keep the fantasy that the text is portable.
- Reference images and cameos. If you still have the stills, those are the identity lock. If identity lived only inside Sora's token, you no longer have it. You will rebuild from frames you export.
- Storyboards. An ordered list of beats is useful. The Sora scene paragraphs are not. Strip them.
Then tag each shot with one destination, using the live catalog rather than a vibe:
| What the old shot actually was | Destination to try first |
|---|---|
| Spoken line, cinematic plate, short | Veo 3.1 |
| Longer take, product or face that must hold | Seedance 2.0 |
| Camera energy, motion, cuts inside one clip | Kling O3 Pro Text to Video |
| One still that should now move | Image-to-video on the row that accepts the plate |
| Several rooms, one character, one film | Movie mode, not a single T2V paste |
If a shot cannot be tagged, it was several shots. Split it. The Sora API sunset migration plan is the deadline. This page is the re-run: how to take a specific old project and not waste a week on a 1:1 fantasy.
Do not start new campaigns on the dying API while you "just finish this board." Finishing means export and rebuild. It does not mean one more Sora batch that you will have to rebuild on the 25th anyway.
Rewrite the prompt for the destination
A Sora prompt is typically a scene narrative: who, where, camera, time, and a soundtrack, sized for 4–20 seconds. Other rows do not share that contract.
Duration. Sora 2 Text to Video listed 4 / 8 / 12 / 16 / 20 seconds. Veo 3.1 lists 4 / 6 / 8. Seedance 2.0 lists 4–15. Kling O3 Pro lists 3–15. A prompt that says "over twenty seconds she crosses the room, picks up the phone, and leaves" is three actions. On Veo it is a crush. Split into one action per generate, or cut in an editor. Do not ask Veo to be 20 seconds by writing "20 seconds" in the prose. The enum on the card is the duration. The prompt cannot extend the card.
Timestamp blocks. A Sora board often looked like [0:00–0:04] kitchen / [0:04–0:12] street. On a model that sells one short clip, that block list is a request for a film. Either you are in Movie mode, or you are writing one shot. Veo can use timestamps inside an 8-second clip to direct beats of the same take. It cannot honour a 40-second board because you pasted the brackets. Seedance and Kling will similarly try to spend the seconds they actually have. Rewrite: one scene, one action, duration set in the form.
Audio in the prompt. Sora had native audio on. A silent Sora prompt still got a track. Veo 3.1 also wants sound written — dialogue in quotes, ambience named — and it is the row that is reliably doing 48 kHz dialogue. Seedance 2.0 generates synced audio in one pass. Kling O3 Pro is audio-true on the Versely card; still write the sound, and still listen. What breaks is a Sora paragraph that asks for a verse of dialogue, a score, and three foley events in a window the new row cannot hold. Cut the line to one sentence. Move the score to a music pass. If the destination is a mute-friendly plate plus TTS, say "no speech" and mean it.
Identity prose. Sora users restated hair, scars, and wardrobe in every scene because the token was sticky and the prose felt like insurance. On Veo reference-to-video and on Seedance, the file is the insurance. Re-describing a face that is already in the reference fights the reference. Rewrite: "She walks the night market. Tracking shot. Lantern light. Audio: wok, crowd." Attach the still.
Cameo and character tokens. They do not exist on the destination. If the old project was "my face in a jetpack," the re-run is a still of that face, then image-to-video, plus whatever likeness and disclosure rules apply on the platform you post to. There is no cameo checkbox to migrate.
Worked rewrite, one kitchen shot, no claim about how pretty the result will be — only about what the form can accept:
Sora-shaped: A 16-second 9:16 of Mara, red wool coat, silver streak in dark hair, setting a ceramic mug on a walnut table, then looking to the window as a bus passes; warm tungsten; she whispers "stay." Fridge hum, bus, spoon.
Veo-shaped: duration 8s on the form. Prompt: Medium shot, Mara in the red wool coat from the reference, sets a ceramic mug on the walnut table. Slow push-in. Warm tungsten. Audio: fridge hum, cup on wood. She says: "Stay." One action. Line short. Coat from the file.
Seedance-shaped: pick a duration the card sells, attach the same still plus a product still if the mug is branded, keep one action. Do not keep "then looking to the window as a bus passes" unless you have seconds for it.
Kling-shaped: lead with the move. Handheld medium, she sets the mug down, camera eases left to the window light. If you needed the whisper more than the move, this was never a Kling job.
Lock identity with stills, not nostalgia
The re-run that works is stills first. Export a hold frame from the old Sora take if the take is the only likeness you have. Generate a clean still on an image row if the hold frame is mush. Sign it off at full size. Then animate.
Image-to-video is the honest port of "this person, now moving." Reference-to-video is the honest port of "this person and this product, in a new scene." Text-to-video is how you start identity retries. It is not how you re-run a project whose whole point was that Mara looked like Mara.
Do not use the old Sora MP4 as a video-to-video restyle and call it a migrate unless you actually want a restyle. Restyle keeps blocking and changes surface. A new model generate from text will not keep Mara. A new generate from her still might.
Brand colours and rooms get the same treatment. One still of the kitchen you already liked is worth more than a paragraph of "walnut, tungsten, late afternoon." Sora ate that paragraph because the board remembered scene one. Veo and Seedance remember what you attach.
If the old project was a six-scene short, lock the character sheet, then either Movie mode or per-shot I2V. Do not paste the six prompts into one 15-second Seedance pass and call the smear a re-run.
Do not expect the seed to travel
A seed is the number that picks the starting noise on that model, with those weights, at that resolution and duration. Same seed, same prompt, same model, same settings — same output. Change the model and the guarantee is void. Seed 42 on Sora has nothing in common with seed 42 on Veo. Neighbouring seeds are not cousins. A provider update on the same row can void yesterday's seed.
So: do not write the old Sora seed into the new job as if it were a colour grade. You can record seeds on the new row once a take is close, and iterate from there. That is a new lock, not a transfer.
The same is true of Sora's other hidden state: storyboard tokens, cameo embeddings, cross-scene colour memory. They were not files in your export zip. If they are not in the help centre's export, they are not coming with you. Rebuild from pixels you hold.
Practical close for one old project this week:
- Export from the sunset URL.
- List shots. Tag each with a live row.
- Pull or make stills for anyone who must recur.
- Rewrite each prompt to one action, one duration the card sells, sound the card can spend.
- Generate. Caption. Cut. Do not wait for the 24th to discover the prompt was still a Sora paragraph.
The AI video generator is the launcher. Image-to-video is the launcher when the still is the point. Open a studio account at Sign up when you have the export and the stills, not when you only have a seed written in a Notion doc.
FAQ
Can I keep the Sora seed if I match the prompt word for word?
No. Seeds do not transfer between models. Matching the prose does not match the noise field. Lock a new seed on the destination after a take is close.
What if I only have the prompt and not the files?
Then you do not have the project. You have a brief. Export is closed or closing; if the sunset page still returns files, take them. Otherwise rebuild from whatever stills you have, and treat the prompt as notes, not as a reproducible job.
Should I re-run every old Sora take on Veo?
Only the ones that were Veo-shaped: short, spoken, cinematic. Product holds go to Seedance. Motion bumpers go to Kling. Films go to Movie mode. A blanket Veo re-run is how you spend dialogue credits on a bottle turn.