Sora is gone

OpenAI announced on 24 March 2026 that Sora was shutting down. The app came off phones on 26 April. The API was switched off on 24 September. Every other story in AI video this year has been an expansion: longer clips, more reference images, higher resolution, cheaper tiers. This is the only contraction, and it belongs to the best-funded company in the category.

No reason was given publicly. What OpenAI published was a help centre notice with dates on it. CNBC framed the closure as cost control. TechCrunch led on the app's reputation. Both are outside readings. The silence is the most useful part of the story, because a company with a defensible reason usually gives it.

The deal that never closed

On 11 December 2025 Disney announced a $1bn investment in OpenAI alongside a three-year character licensing deal attached to Sora. It did not close. No money changed hands. Reporting on the API shutdown treats the two as a single event, which is fair. The licensing was the commercial case for a consumer video app, and when the app went, the case went with it.

For anyone planning brand work, this is the instructive half. The largest licensing arrangement proposed in this field, by the studio with the most valuable character library in the world, lending it to a generative model for three years, was announced, reported everywhere, and quietly did not happen. A licensing announcement in AI video is not yet evidence of anything. The question to ask is whether money has moved.

The economics nobody puts in a deck

Some of the shape of this is visible from public pricing elsewhere. Artificial Analysis publishes cost per minute of generated video across the frontier models. ByteDance's Seedance 2.5 sits at $34.12 a minute. The previous version is $22.45. The cheapest frontier option on the board, MiniMax H3 at 768p, is $4.80. The figures come from the same board that ranks quality.

Seedance 2.5

$34.12

Previous version

$22.45

MiniMax H3

$4.80

Now put a free feed of generated video in front of several million people. Every scroll is a generation. There is no cached asset to serve twice, because the entire proposition is that the next clip has never existed. Cost rises with engagement rather than being spread across it. Nobody outside OpenAI knows its unit economics, and we are not pretending to. What is observable is that video is the only consumer AI product where success and cost point in the same direction, and that the one lab which tried it at the largest scale stopped.

“A lot of what the industry currently believes about AI advertising quality rests on output from a model nobody can run.”

The second-order point is about forecasting. Every chart in every deck on this subject in 2026 points up, because until March there was no reason to separate two different axes: what a model can do, and whether it will still be there. Capability has kept climbing. Availability now has a precedent going the other way.

If you built a pipeline on it

Six months from announcement to the API going dark. One month for the app. That is the real planning number, and it is shorter than the lead time on a lot of annual campaign work.

Two things that got stranded. The Toys"R"Us origin film, made by Native Foreign with Toys"R"Us Studios and premiered at Cannes in June 2024, was one of the first commercials produced with Sora. It took about twelve people roughly three months, which NBC News noted was roughly the same as any other job. The film still exists. The thing that made it cannot be run, so nothing in it can be extended, re-cut with new shots, or matched for a sequel. A brand asset built that way has a hard edge around it from the day the model closes.

The other is research. On 18 May 2026, Ipsos and Syracuse University published a study on AI-made advertising whose test ads were generated with Sora 2. The paper outlived the model by four months. Its finding, that AI ads are already good enough and that this is the problem, may hold up perfectly well. But anyone who wants to replicate it or argue with it cannot reach the generator that produced the stimulus. A lot of what the industry currently believes about AI advertising quality rests on output from a model nobody can run.

What we do about it

A model is a camera body, not a studio. Chosen per job, named in the treatment, and replaceable. That sounds obvious until you look at what a production actually accumulates while a project runs, and sort it into two piles.

• Survives a shutdown. The shot list, the boards, approved stills, reference frames, the grade, the sound design, the written prompt record. All of it describes intent rather than a particular model's behaviour, and all of it transfers.
• Does not survive. A character that exists only as a fine-tune. A look that one model produces and no other matches. An asset library you cannot re-render. Anything the brand has to keep making.

Which gives one question worth asking at briefing, before the model is chosen: if this model is gone in ninety days, what part of this campaign stops? If the answer includes something the client needs to produce again next quarter, either pick differently or build the fallback while the budget is still open.

The practical habit is smaller than it sounds. For every shot we keep the final frame as a still, the full prompt and parameter record, and the reference images in the state they were fed. It is three or four hours of admin on a project and it means a film can be rebuilt, by hand or on another model, if it has to be. We were doing some of this already for continuity. Sora turned it into the reason.