Sora 2 Shuts Down September 24, 2026: Migration Guide and Alternatives
OpenAI removes the Sora 2 models and Videos API on September 24, 2026, with no replacement model. Here is the timeline, what to export, and which models to migrate to by use case.
OpenAI is retiring Sora. The consumer apps are already gone, and the API follows on September 24, 2026. Unusually for a deprecation of this size, OpenAI's own documentation lists no recommended replacement model — this is an exit from the product line, not an upgrade path. Teams built on Sora 2 need a different vendor, not a new model string.
The Timeline
| Date | What happens |
|---|---|
| March 24, 2026 | OpenAI notifies developers of the deprecation |
| April 26, 2026 | Sora web and mobile apps discontinued |
| September 24, 2026 | Videos API and all Sora 2 models removed; calls return 410 Gone |
The removal covers every alias and snapshot: sora-2, sora-2-pro, and their dated versions, along with the Videos API endpoint itself. After the cutoff, OpenAI permanently deletes data associated with Sora use, with no recovery window.
Export Your Work First
If you generated anything in Sora that you still need, export it before the date rather than after. OpenAI provides an export flow at sora.chatgpt.com/sunset, and recommends doing it as soon as possible rather than waiting for a final window. Anything not pulled onto your own storage is gone when the cutoff hits.
What To Migrate To, By Use Case
Sora 2 was a generalist, so there is no single one-for-one swap. The practical approach is to match each thing you were using it for to the model that does that specific job best.
Cinematic, photorealistic motion
Kling 3.0 is the closest match for the work most people used Sora 2 for: dramatic camera movement and physically plausible motion in a polished frame. It runs 5s and 10s at 720p or 1080p, from 65 credits, with audio available as an option. Kling 3.0 Turbo trades a little quality for faster turnaround.
Long shots and scenes with dialogue
Seedance 2.5 goes well past what Sora 2 offered on length: up to 30 seconds in a single pass, with native audio generated alongside the picture at no extra credit cost. It tops out at 720p, so if you need 1080p, pair it with Kling 3.0. Full details in the Seedance 2.5 guide.
Animating a still image
Hailuo 2.3 is the cheapest way to bring a photo to life, starting at 25 credits for a 768P 6-second clip. Hailuo 2.3 Pro handles the same job with more motion fidelity. Both are image-to-video only, which is exactly the constraint you want when that is the whole task.
Keeping a character or product consistent
This is where the newer models clearly exceed what Sora 2 could do. Seedance 2.5 accepts up to 9 reference images and Wan 2.7 accepts 5, teaching the model what a subject looks like before it animates anything. For brand and character work this is the difference between usable output and a face that changes between shots.
Restyling footage you already have
Kling 3.0 Motion Control takes an existing video and transfers its motion, which was never something Sora 2 handled well.
Side-By-Side
| Model | Max duration | Resolutions | From | Strength |
|---|---|---|---|---|
| Kling 3.0 | 10s | 720p, 1080p | 65 credits | Cinematic motion, optional audio |
| Kling 3.0 Turbo | 10s | 720p, 1080p | 85 credits | Faster iteration |
| Seedance 2.5 | 30s | 480p, 720p | 80 credits | Long single takes, audio included |
| Seedance 2 Mini | 10s | 480p, 720p | 45 credits | Cheap drafts, 9 references |
| Wan 2.7 | 10s | 720p, 1080p | 75 credits | Reference control at 1080p |
| Hailuo 2.3 | 10s | 768P, 1080P | 25 credits | Cheapest image-to-video |
A Migration Checklist
- Export before September 24. Treat the date as hard, because 410 Gone is not negotiable.
- Inventory what you actually used Sora 2 for. Most teams find two or three real use cases, not one.
- Re-test your prompts. Prompt phrasing does not transfer cleanly between vendors. Budget a round of cheap low-resolution tests rather than assuming your Sora prompts will work verbatim.
- Avoid rebuilding the same dependency. The lesson of this shutdown is that a single-vendor video pipeline is fragile. Wiring your workflow to two models with different providers costs little and removes the single point of failure.
Start Testing
All the models above are available now in Text to Video and Image to Video, with reference-driven generation in Reference to Video. New accounts include free credits, so you can compare candidates against your own prompts before the deadline rather than after it.