Sora is gone: the risk of relying on a single AI model for a brand film

On March 24, 2026, OpenAI announced it was shutting down Sora, the AI video app it had launched about six months earlier. The same day, it told developers that the video API and the Sora 2 models would go offline on September 24, 2026.

That date has passed. For anyone who makes or commissions AI brand films, the case matters less as news and more as a risk test: what happens to a project when its central tool disappears?

What happened, with dates

• March 24. OpenAI announces the end of the app. According to TechCrunch, the company gave no reason and no date that day.
• March 24. OpenAI’s official deprecations page lists the end of the Videos API and the sora-2 and sora-2-pro models, dated September 24. The recommended replacement column is empty.
• April 26. According to OpenAI’s help center, the app and the web version were discontinued. The page also says user data will be deleted after discontinuation and recommends exporting content.
• September 24. The scheduled end of the API.

TechCrunch also reported that the deal with Disney, which would have allowed videos with Marvel, Pixar and Star Wars characters in Sora, collapsed with it. According to the report, it appears no money changed hands.

The app’s numbers

TechCrunch cited Appfigures estimates: about 3.33 million downloads at the peak, in November, and 1.13 million in February. Lifetime in-app purchase revenue came to around US$ 2.1 million. These are estimates from an analytics firm, not OpenAI data.

On March 29, TechCrunch published an analysis of the episode. Its authors read Sora’s exit, together with ByteDance’s delay of the global launch of Seedance 2.0, as a sign that claims AI would replace film production were overstated, given the technical and legal obstacles. That is the publication’s reading, not a measured fact.

Why it matters for a brand film

An AI brand film usually starts from one tool: the look of the shots, a character’s face, the camera movement are all learned on top of a specific model. When the model changes or disappears, what was “taught” to it does not come along.

There are four concrete risks:

1. Continuity. A campaign with several films over the year needs a stable look. Switching models midway breaks the identity of characters and light.
2. Schedule. Redoing a shot in another tool costs days, not minutes, because each model answers the same text differently.
3. Files. What lives only inside the platform can be deleted. OpenAI itself warned it would delete Sora data.
4. Contract. The client paid for a film, not for a model. Whoever delivers carries the risk of the tool changing.

What to do in practice

None of this requires giving up AI. It requires treating the model as a supplier that can leave.

• Keep everything outside the platform. Final shots, reference images, character sheets and prompts live in your own versioned files.
• Write the film in text and image, not in “the model’s way”. Script, shot list, palette and character sheets work in any tool. A prompt pasted from a single platform does not.
• Have a second model tested. Before you need it, run a test shot in another tool and note what changes.
• Ask for the end-of-life date. Before closing a long campaign, check the supplier’s notices page. OpenAI gave six months’ notice; not every supplier will.
• Agree in the contract on what the client receives. The final files and references, not access to the tool.

This blog’s point of view

Sora was not a bad model that lost to another. It was a product the company decided to close. That can happen to any tool, from any maker, for reasons the client will never know.

The defense is process, not brand choice. It is the same logic as the model-by-stage map: no model does everything, and none should be the only one. In AI brand films, continuity of the result must stay with whoever directs, not with the platform.

This text was written with AI support, from the official and press sources cited above. The cover image was generated with AI.