On 24 September, OpenAI turns off the Sora API. The consumer app went dark on 26 April, a month after the company announced the wind-down on 25 March, and the help-centre notice reads like a decommissioning order: export your content, because everything associated with your use of Sora will be permanently deleted. The product that made text-to-video a household idea in February 2024 had, by the time it was cancelled, fallen from first to 126th on the App Store's free chart, and Disney had walked away from a three-year licensing deal (announced barely three months earlier, and reported to include a planned $1bn stake) that would have let more than 200 of its characters be generated on demand.

Six weeks before that API shutdown date, Higgsfield closed a $400m Series B at a $5.4bn valuation, roughly four times the $1.3bn it was worth in January, on $700m of annualised revenue, 30 million users across 238 countries and 390 Fortune 500 customers. Both of these things are true, and neither is a contradiction. The consumer destination for generated video is closing. The industrial supply of it is compounding.

What has actually happened is that generative video stopped being somewhere you go and became a stage in a pipeline. For anyone who builds physical experiences, that reclassification is the whole story.

The arithmetic of a minute

Put the two ends of the market next to each other.

At the model end: Veo 3.1 generates in eight-second clips at 1080p and 4K, chainable to something over 140 seconds, and is listed by API resellers at roughly $0.40 per second with synchronised audio, $0.20 without. Call it $24 of raw generation for a minute of moving image with sound. A decade ago that minute was a shoot.

At the venue end: The Wizard of Oz at Sphere runs 75 minutes on a 160,000-square-foot, 16K-by-16K screen, built by taking a 1939 Technicolor film in a four-by-three frame and using custom-tuned Veo, Imagen and Gemini models for super-resolution, outpainting the edges of shots that were never composed to be seen whole, and generating performances for actors the original camera had panned away from. Around 1.2 petabytes of data went through it. Thousands of researchers, engineers, VFX artists, archivists and producers were involved across Google DeepMind, Google Cloud, Sphere Studios, Magnopus and Warner Bros. Discovery. The New York Post reported a budget of $80m in 2024, when the show was still an 80-minute plan; later accounts put it nearer $100m. Against the 75 minutes that actually opened, that is somewhere north of a million dollars a minute.

Same technology family. On the $80m figure, about forty-four thousand times the cost per minute.

It is not a like-for-like comparison ($24 buys model inference, $80m buys an entire production), and the mismatch is the thing worth measuring. The gap is a precise inventory of everything the model does not do: hold a 16K frame together across a curved surface, survive colour management across thousands of LED panels, sit on the right side of a rights agreement with Warner Bros., and still be there, frame-accurate, on the eleventh show of the week. Buzz Hayes, Google Cloud's global lead for entertainment industry solutions, described the work as "an epic production undertaking, achieved with still new and largely experimental technology". That is an unusually honest way of saying the generation was the part they could buy.

The model is now the cheapest thing in the room. Everything that makes its output survive contact with a real wall is not.

The feed is demoting it. The box office is paying for it.

The consumer numbers explain why the app model broke. Animoto's January report (450-plus respondents, surveyed the previous September) found 83% of consumers say they have watched a video they suspected was AI-generated, and 36% say AI-generated video lowers their perception of the brand behind it; 68% want to see real people. The platforms have read the same data. Since July, Snapchat no longer recommends wholly AI-generated videos in Spotlight. That is a stricter line than TikTok, YouTube or Meta, which mostly label and deprioritise rather than exclude.

Meanwhile the advertising middle is quietly saturated. IAB's 2025 video research found a majority of ad buyers already using generative AI for video creation, 30% of digital video ads built from scratch or enhanced with it, and a forecast of 39% for 2026; its separate July 2026 report puts US digital video ad spend past $80bn. Generated video is not a differentiator in that market. It is a cost floor.

The Sphere show, by contrast, has sold more than three million tickets and taken over $400m since it opened in August 2025. Nobody bought those tickets to watch AI. They bought them to be inside something at a scale that cannot be reproduced on a phone. Generative models were how a 1939 negative was made big enough to fill it.

What a generated frame actually costs on a wall

At BCVR we have watched the same line-item collapse everyone else has. Concepting, previs, texture and plate work that used to consume weeks of a schedule now resolve in an afternoon, and on some projects the raw generation cost has fallen far enough that it no longer appears as its own budget line.

Almost none of that saving reaches the bottom of the sheet, because it was never where the cost lived. The money on a hybrid physical-digital build goes into warping and blending across projector or LED batches that never quite match; colour-managing generated content that was trained on consumer footage and arrives with the wrong gamma; cutting loop seams that hold up on the four-hundredth pass in front of a queue; provenance paperwork for every generated asset in a client's brand-safety review; and the on-site week when someone has to look at the actual wall and say it is not right yet. What generative models changed is not the price of the room. It is how many ideas a studio can afford to test before committing to one, and that shows up, when we measure it, as dwell time and throughput rather than as render savings.

Buy the pipeline, not the prompt

The lesson of the Sora sunset is not that generative video is overrated. It is a governance point with a date on it: a capability many teams treated as permanent infrastructure had a discontinuation notice attached, and anyone whose workflow assumed that specific endpoint has one month to move. Model-agnostic pipelines, local copies of what you generated, and provenance records that survive a vendor's strategic pivot are no longer good hygiene. They are the difference between an asset library and a hostage.

For studios and commissioners the practical read is simpler still. Generation is now the commodity input; the defensible work is everything downstream of it, in the physical envelope where the audience actually stands. Price the room, not the render.

Scott Marshall is founder of BCVR (https://www.bcvr.io), a London-based immersive production studio working across XR, projection mapping, spatial audio and digital twins for brand and cultural clients.

Sources

  1. OpenAI Help Center, "What to know about the Sora discontinuation": app discontinued 26 April 2026; API discontinued 24 September 2026
  2. eMarketer, "OpenAI to discontinue Sora app as Disney exits partnership" (announced 25 March 2026; App Store #1 to #126; Disney three-year deal, planned $1bn stake, 200+ characters)
  3. Wikipedia, "Sora (text-to-video model)": release and discontinuation timeline
  4. PR Newswire, "Higgsfield Raises $400 Million Series B Financing at $5.4 Billion Valuation…", 17 August 2026
  5. Google DeepMind, Veo model page: 1080p and 4K output, eight-second generations, scene extension
  6. OpenRouter, Veo 3.1 listing: $0.40/second with audio, $0.20/second video-only; up to 20 chained clips for 140+ second narratives
  7. Google, "How Sphere's new 'The Wizard of Oz' experience is coming to life with AI"
  8. Sphere, "The Wizard of Oz at Sphere: AI" (160,000 sq ft, 16K x 16K)
  9. TVBEurope, "How Google used AI to 'reconceptualise' The Wizard of Oz": Buzz Hayes quote
  10. Sphere Entertainment Co. investor release, 16 June 2026: 3m+ tickets, $400m+ ticket sales since 28 August 2025
  11. Las Vegas Review-Journal / Neon, 5 August 2024, reporting an unconfirmed New York Post figure of $80m against an 80-minute plan
  12. Wikipedia, "The Wizard of Oz at Sphere": 75-minute runtime, ~1.2 petabytes, $100m budget figure
  13. Business Wire, Animoto 2026 State of Video Report, 21 January 2026 (450+ respondents, surveyed September 2025)
  14. Marketing Dive, 18 July 2025, reporting IAB's 2025 Digital Video Ad Spend & Strategy Report (Part Two): 30% of digital video ads, forecast 39% for 2026
  15. IAB, 2026 Digital Video Ad Spend & Strategy Report, 14 July 2026: US digital video ad spend surpasses $80bn in 2026
  16. MLQ News, "Snapchat cuts fully AI-generated videos from Spotlight recommendations", July 2026