The Sora API Shuts Down Today. Your Finished Videos Survive. Your Workflow Doesn't.
What actually breaks when a hosted video model retires, and the one-hour audit that tells you how exposed the rest of your work is.
OpenAI's help page about the Sora shutdown answers five questions, and four of them are about getting your files out: whether you can export, how to export, what happens to your data afterwards, how to get a refund. The remaining one is the timing, and it is two sentences long: "The Sora web and app experiences were discontinued on April 26, 2026. The Sora API will be discontinued on September 24, 2026."
That is today. And the page is a good page, clear and early and honest, with a link to the export button and a plain recommendation to use it soon. It just answers the wrong question. Because for most people who were actually using Sora to make things, the footage was never the fragile part.
The part the shutdown notice doesn't cover
Here is the sentence I have been thinking about all week, and it did not come from OpenAI. It came from ElevenLabs, in a changelog entry dated yesterday, September 23, explaining what the Sora retirement means for people using Sora inside their platform:
Flows and templates that use a Sora node need to be switched to another video model, such as Gemini Omni 1.1 Flash, before they can run again.
The same entry confirms your existing generations stay in your history. So the clips are safe. The thing that is not safe is the assembly. If you built a repeatable process, the sequence you run every time a client sends a brief, the template you duplicate for each episode, the chain of steps that turns a script into six social cuts, and a Sora call sat anywhere in that chain, the whole thing stopped working today. Not degraded. Stopped.
This is the difference between a file and a recipe, and almost nobody plans for it. You back up footage. You do not back up a process, because a process does not feel like a thing you own. It feels like knowledge. Then one link in it turns out to have been rented.
The retirements are not a one-off, either. That same ElevenLabs entry dates a second one: ByteDance retires Seedance 1.5 Pro on November 11, 2026, and it is already unavailable for new generations. Two video models with death certificates in the same paragraph.
The replacement is narrower than it looks
The obvious response is "fine, swap the node." That is the right instinct, and if you are in the United States it is worse than it sounds.
I went through ElevenLabs' full list of video models to see what a US-based person actually has to swap to. The Seedance family is there in six variants, and every single one of them carries the same line: "Seedance 2.0 is not available in the United States." Same for Fast, Mini, 2.5, 2.5 Video Edit, 2.5 Video Extend. The Kling family, three entries, carries the same line. Google's Veo 3.1 variants carry no such restriction.
So for a US creator, the practical replacement set is not "a dozen video models." It is Google's and Gemini's, and that is a real constraint on the work, not a branding question. Compare what you could ask for:
| Seedance 2.0 | Veo 3.1 | |
|---|---|---|
| Length | 4 to 15 seconds | fixed at 4, 6, or 8 seconds |
| Aspect ratios | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 | 16:9, 9:16 |
| Resolutions | 480p, 720p, 1080p | 720p, 1080p |
| Available in the US | no | yes |
If your dead Sora shot was a twelve-second cinematic wide, there is nothing on the US-available side of that table that matches it. You are not swapping a node. You are re-cutting the shot, or re-briefing the client, or stitching two eight-second pieces and hoping the seam hides.
One more thing buried in the same documentation that is worth knowing before you plan a migration: ByteDance models are disabled by default for API requests and need explicit approval. And for enterprise workspaces, every image and video model is disabled by default until an admin turns it on. So "the model exists on this platform" and "I can call it this afternoon" are two different facts, and the gap between them is a support ticket.
What I actually think is happening
Hosted models are better than what most of us can run at home. They are faster, they look nicer, and they cost less than the electricity plus the hardware plus the afternoon. That is not going to change, and anyone telling you to move everything onto your own machine is selling you a large tax to insure against a small risk.
But the risk is no longer theoretical, and it has a specific shape. It is not "the model gets worse." It is "the model gets a date." Two got dates this week. A third thing happened on Wednesday that rhymes: Adobe finished buying Topaz Labs, the upscaler a lot of photographers and editors bought outright years ago. Adobe said the right things in public, in writing, including that the Topaz apps "continue to be available as standalone products" and that the brand stays. I believe them. It is still true that an app you owned is now a line item inside a subscription company, and the on-device processing path that made it useful to people who work offline is now somebody else's annual strategy decision rather than a property of the software on your drive.
None of these three events is a scandal. Put together, they are a straightforward argument that the question worth asking about any tool in your process is no longer "is this the best one" but "what happens to my finished work if this disappears, and how long would it take me to find out?"
Put this into practice
The export takes two minutes. The audit takes about an hour and is the part worth doing.
1. Export your Sora content right now, before you read the rest of this. Go to sora.chatgpt.com/sunset and click Export. You get an email when the archive is ready, so it is not instant, which is exactly why you start it before you need it. OpenAI's help page says that after any final export window passes, it "will permanently delete any data associated with your use of Sora." One detail people miss: purchased ChatGPT and Sora credits are not stranded, they can still be used for Codex.
2. Open your five most-repeated jobs and list them by name. Not your favourite experiments. The ones you actually run for money or on a schedule: the weekly short, the client thumbnail set, the product clip template, the podcast cutdowns. If you have templates or saved flows, those are the list.
3. For each one, write down every model it calls that you do not have a copy of on your own drive. API nodes, cloud services, anything with a login. Be pedantic. Include the upscaler, the background remover, the voice, the music. That list is your exposure and it is usually longer than people guess.
4. Next to each model, write one word: rebuild or abandon. This is the whole exercise. Decide now, while nothing is on fire, what you would do if that name went dark next Tuesday. Some of them you would happily drop. Some of them are load-bearing for a paying client, and those are the ones where you want a tested alternative identified before you need it, not during.
5. For anything marked rebuild, run the alternative once this month on a real job. Not a test prompt. A real deliverable, at the real length and real aspect ratio, so you find out about the fixed-duration problem now rather than on a deadline. Write down what the substitute costs and what it can't do, in a file next to the workflow. That note is the actual deliverable of this whole exercise.
If you want the ten-minute version instead of the hour: do steps 1 and 3 only. Knowing your exposure without having decided what to do about it still beats not knowing.
Honest limitations
This is an audit, not a solution, and I want to be clear about what it does not fix.
It does not protect you from a price change, which is the far more likely event and much harder to plan for than a shutdown. A model that triples in cost is still available and your process still runs; you just cannot afford it. Nothing in the exercise above catches that.
It also does not make the local alternative safe. The open-weights releases that landed this week are governed by real licences with real restrictions. The official ComfyUI packaging of Qwen-Image-2.1, the most useful open image model most people can run right now, is published under qwen-research, pointing at Qwen's own licence terms. Black Forest Labs' new FLUX 3 Action ships under something called the FLUX Kommunity License. Weights on your disk cannot be removed from a dropdown, and that is genuinely worth something. They can still carry terms that stop you selling what you made with them, which is a different failure with the same effect on your invoice.
And the audit is boring. It produces no images. It will not make anything you can post. I have skipped it myself more than once and the only reason I am writing about it today is that I watched a shutdown notice turn into other people's broken templates in real time this week, and it took about four sentences of a changelog to see it coming.
What I'd watch next
The thing I do not know is whether any of this gets better. Hosted model retirement is still being handled as a product decision announced on a help page, and the tooling around it is a per-platform courtesy: ElevenLabs wrote a clear changelog entry naming the exact breakage, which is more than most will do. There is no standard here. No deprecation window anyone agreed to, no machine-readable way for a workflow to warn you that one of its steps has a sunset date, no convention that a retired model gets a documented equivalent.
That seems fixable, and cheap, and nobody is doing it. If your own tools have a way to flag a dependency that is about to expire, I would like to know what it is, because I went looking this week and found four vendors doing four different things and one of them doing nothing at all.
In the meantime the only version control you have for a process is the one you write down yourself.
Medium metadata
- Title: The Sora API Shuts Down Today. Your Finished Videos Survive. Your Workflow Doesn't.
- Subtitle: What actually breaks when a hosted video model retires, and the one-hour audit that tells you how exposed the rest of your work is.
- Tags: AI Video, Sora, Creative Workflow, Video Editing, Generative AI
- Primary sources: OpenAI Sora discontinuation FAQ · ElevenLabs changelog, September 23, 2026 · ElevenLabs image and video model list · Adobe on completing the Topaz Labs acquisition