Die Sora-API wurde am 24. September abgeschaltet. Was Marken daraus für KI-Videos lernen können

On September 24, 2026, OpenAI removed its Sora video models from its API. It had announced the change six months earlier, on March 24, and the Sora app and website had already closed on April 26. As of October 3, the replacement column on OpenAI's own deprecations page still holds a dash where a successor model would normally be named. A brand building on the API got six months of notice. A brand working in the app got about a month.

Sora was not one of the four engines in our production rotation, so this is not a post-mortem of our own. The question it raises still applies to all four, and one of them, Veo, is changing on Google's own API this month, as the list below shows. We think every brand buying AI video should put that question to its studio or its tool: if the model that made your ad disappeared tomorrow, what would you lose?

What happened, in dates

• March 24, 2026: OpenAI posted that it was saying goodbye to Sora and promised timelines for the app and the API, plus details on preserving users' work. Its deprecations page listed the Videos API, sora-2 and sora-2-pro, with their dated snapshots.
• April 26, 2026: the Sora app and web experience closed. OpenAI's help page says exports remain available after that closure, recommends exporting as soon as possible and says it will notify users if a final export window is offered.
• September 24, 2026: the Videos API and every listed Sora model were removed. The oldest snapshots on the list carry the date 2025-10-06 in their names, less than a year before the removal.
• Replacement listed by OpenAI: none.

Why it closed, as reported

OpenAI's announcement did not give a detailed reason. NBC News, reporting the same day and citing The Wall Street Journal, said OpenAI's top executives had been talking about sharpening the company's focus, and noted that moving computing power away from Sora could free chips for more lucrative coding, reasoning and text work. NBC also reported, citing a source familiar with the matter, that Disney's three-year deal to bring its characters to Sora, announced in December 2025 with a planned $1 billion investment in OpenAI, was not proceeding.

For a brand, the reason matters less than the pattern. A model is a product. Products are withdrawn for reasons that have nothing to do with how good they are, or with how much a customer relies on them.

2026 has been a year of arrivals and exits

Sora is the loudest example this year, not the only one:

• February: the Motion Picture Association sent ByteDance a cease-and-desist over its AI models. Engadget later noted reports that ByteDance suspended the global rollout of Seedance 2.0.
• June 16: xAI made Grok Imagine Video 1.5 generally available on its API.
• June 30: Google retired its Veo 3.0 models from the Gemini API.
• July 31: ByteDance's Seed team launched Seedance 2.5 and doubled a single generation from 15 to 30 seconds.
• August 17: the Motion Picture Association announced a memorandum of understanding with ByteDance on copyright guardrails for AI video and image models. Its announcement names Seedance 2.5 among the recent launches.
• August 27: Google released Gemini Omni 1.1 Flash, which now replaces Veo 3.1 for video in the Gemini app.
• September 28: Kling opened limited early access to Kling 4.0 Flash. As of October 3, Kling's full 4.0 feature page still says Coming Soon in October; early access is not the full release.
• October 1: Runway, which hosts several of these models, announced it is available inside OpenAI's new always-on agents, called Dots.
• October 22, scheduled: Google retires its Veo 3.1 preview models from the Gemini API and names Gemini Omni 1.1 Flash as their replacement.

Some are upgrades, one was a pause, and some are exits. Each changes what a given prompt produces. A new version is good news and a small migration at once, because the same prompt on a new model is a different shot.

What actually breaks when a model disappears

The finished file does not break. An exported film plays the same on September 25 as it did on September 23. What breaks is everything you might want to do with it next.

• The revision. A note on one shot means making that shot again. If its model is gone, the new shot comes from another engine and has to be matched by eye to the shots around it.
• The character. A recurring face that exists only as a saved character inside one platform ends when the platform does.
• The look. Texture, color and motion differ from engine to engine. A campaign whose look was simply whatever one model did has no way back to it.
• The next version. A new language, a vertical edit, a holiday variant of last quarter's ad. Each is straightforward if the parts live outside the model, and a rebuild if they do not.

Own the film, rent the engine

In our script-led ads, the film exists before any engine runs. You approve the script and the sound before any footage is made. Then the engine is chosen shot by shot from the four we run (Seedance, Kling, Veo and Grok Imagine), depending on what each shot needs. A finished 30-second ad might carry shots from three engines, and the viewer sees an edit, not a model.

That habit grew out of quality, not fear of shutdowns, but it is what makes a shutdown survivable. If one engine went away tomorrow, the script, the approved sound, the edit and the type would all still exist outside it. The shots it made would need a new home. The film would not need a new idea.

Three working rules carry most of the weight, with any vendor:

• Keep words out of the generation. Generated on-screen text tends to garble, so we generate a clean plate and set the type in the edit. A side effect: a new price or a translated line on screen is a new type layer, not a new generation.
• Keep references in a form more than one engine accepts. Engines differ here. In our work, Seedance has refused AI-generated images as reference inputs and wanted real photographs. A product reference built from the client's own photographs travels further than a still only one engine will take.
• Keep the sound separate. A voice and music bed approved before the picture can carry a replacement shot. Audio generated together with a clip lives inside that clip, and when the shot is replaced, its sound goes with it.

“Own the film. Rent the engine.”

Six questions to ask any AI video vendor

• Which engines made my ad, and which shots came from which?
• If one of those engines is withdrawn, how would you handle a revision to its shots?
• What do I receive at delivery: final exports only, or project files too?
• Where do the approved script, the sound and the reference images live, and can I have copies?
• Could you make a new language or format version of this ad a year from now, and what would that need from me?
• Who holds the usage rights to the finished film, and does anything in them depend on one engine's terms?

Our own answers to two of them are already on the site. Every plan delivers the final exports, and if you also need the project files, say so in your brief and we will tell you what we can hand over. You own every finished film we make for you: all rights transfer to you on delivery.

Keep your masters

OpenAI's help page says users can export their work after the app's closure and recommends doing so as soon as possible. It says that once the service and any final export window have ended, the data tied to their use of Sora will be permanently deleted. A platform library is a convenience, not an archive. Whoever makes your ads, keep these somewhere you control:

• The final exports, in every format and language you ran.
• The approved script and the sound: voice, music and the final mix.
• The reference pack: product photographs, approved stills and character sheets.
• A short note of which engine made which shot, and when. Models change monthly, and the note tells the next person what they are matching.
• The usage terms and the delivery confirmation.

We say this as a studio that learned part of it the slow way. Our delivery report counts 999 client projects and 9,721 finished files from 2017 to 2026, which we could only do because the delivery folders still exist. It has no year-by-year chart, because a storage migration in 2021 left earlier file timestamps unreliable. The files survived the move. Some facts about them did not. And the 742 projects we could read cut by cut were readable because their files were named 1.0, 2.0, 3.0. Put the facts that matter into names and notes, not only into the system that happens to hold them today. The full report is at filmito.io/en/ai-ads-report.

If you were building on Sora

There is no one-for-one swap, and we will not rank tools here. Start from the shot, not the model: which shots need a face that holds across a scene, which begin from an approved still, which can be described tightly in a prompt. Our notes on how we hand shots to Seedance, Kling, Veo and Grok are at filmito.io/en/blog/seedance-vs-kling-vs-veo-vs-grok-for-ads. They describe late August, and the engines have moved since.

A note on agents

The next layer is arriving. On October 1, Runway said it was available inside OpenAI's always-on agents, called Dots, pitching a brief given at bedtime and a finished spot waiting in the morning. That may be handy for drafts. It also puts one more product between a brand and its film, and products can be withdrawn. The rule stays the same: own the film, rent the engine.

If you want an ad built so that no single model decides its future, filmito.io/en/see-your-ad asks six short questions and a person replies within one business day. What each package includes is at filmito.io/en/pricing.