On September 24, 2026, the last working piece of Sora goes dark. OpenAI pulled the consumer app and sora.com back in April, and the API has run on borrowed time since. After this week there is no legitimate way to generate a Sora video at all. That is the quiet ending for a product that, ten months earlier, sat at number one on the App Store.
I wrote about the OpenAI Sora shutdown when it happened in the spring, and the argument was simple: OpenAI did not retire Sora because the videos were bad. It retired Sora because the arithmetic never worked, and no amount of scale, pricing, or optimization was going to close the gap. Five months of hindsight is enough to grade that call. Most of it held. The part that did not is more useful than the part that did, because it points at where the whole industry is heading next.
The number everyone quoted, and the one that held up
The original post led with a brutal figure: roughly $130 in compute per ten-second clip, an estimated $15 million a day at peak usage, against $2.1 million in total lifetime revenue. That $15 million number went viral because it was staggering. It also deserves an asterisk that newer reporting has since attached.
Later accounts, including Wikipedia’s timeline of the model, put Sora’s steady-state operating cost closer to $1 million a day rather than $15 million. The higher figure appears to have been a peak-load estimate from the busiest weeks; the lower one reflects day-to-day running costs after usage collapsed. The honest way to state it now: the real number sits somewhere in that range, and it does not matter which end you pick. Even the conservative $1 million a day, annualized, is roughly $365 million against $2.1 million earned across the product’s entire life. That is a business losing something like 99 cents on every dollar of revenue. The correction to the headline does not soften the conclusion; it sharpens it.
The rest of the timeline has firmed up too. OpenAI declared the free tier “completely unsustainable” 32 days after launch and bolted on paid generation in November 2025. The Disney licensing deal, announced December 11, 2025, put $1 billion on the table for a three-year partnership across Disney, Marvel, Pixar, and Star Wars characters. Sora fell from the top of the App Store to number 165 by March 2026, downloads dropped 66% from their November peak, and OpenAI announced the shutdown on March 24. Disney reportedly learned its billion-dollar partner was gone less than an hour before the public announcement.
What the call got right: video moved to where the money is
The strongest claim in the original piece was not “Sora failed.” It was narrower and more testable: consumer-priced AI video does not work at current compute costs, and the technology would survive only in contexts where customers pay enough to cover the bill. Five months later, that is exactly the shape the market took.
Video generation did not die. It moved upmarket. The field that serious creators actually argue about now, which I covered in depth in the AI video generator comparison, is a set of priced-for-cost professional tools: Google’s Veo 3.1 for synced-audio cinematics, Kling 3 for volume production, Runway for editing workflows, ByteDance’s Seedance for quality per dollar, and, tellingly, a surviving paid Sora 2 Pro tier that ran at roughly $0.10 to $0.70 per second depending on resolution. The price spread across that frontier is now about sevenfold, from a dime a second to seventy-five cents.
Notice what died and what lived. The free, subsidized consumer firehose died. The tier where a studio pays enough per clip to cover the compute lived. That is the thesis, confirmed almost to the letter: willingness to pay met the cost of inference only at the professional end, and the professional end is where AI video quietly became a real business while the consumer app was making headlines for the wrong reasons.
Where the call was too clean: the enterprise pivot is not a bottomless well
Here is the part I got too neat. The spring post framed the enterprise pivot as a safe harbor. Compute freed from rendering cat videos would flow to ChatGPT Enterprise, API inference, and coding agents, where revenue comfortably exceeds cost. The implication was that “enterprise” is a category where the economics simply work.
The last five months complicated that. Enterprise AI spending is now getting squeezed by the same force that killed Sora. Ramp’s August 2026 AI Index, drawn from roughly 70,000 companies, found that AI spend per employee among the heaviest adopters fell 9.7% in a single month, to $7,205 from $7,976 in July. Effective token prices dropped to $0.68 per million, down from $1.15 in March, a 41% slide in half a year. The share of tokens going to premium frontier models fell from 53% to 45% as buyers migrated workloads down to cheaper standard models.
That does not mean enterprise AI is failing. Paid adoption actually ticked up. It means “revenue exceeds cost” is not a fixed property of the enterprise tier the way I implied. The compute-economics discipline that OpenAI enforced on Sora is now being enforced on the labs themselves, by their own customers. When GPT-6 Astra launched in September at $10 per million input and $50 per million output, buyers pushed back on price and routed volume to cheaper models, a dynamic I traced in why the token-spending era is ending and in the growing model-fatigue problem of picking a model at all. The pivot to enterprise happened. It just is not the infinite runway the spring framing assumed.
The compute-economics story stopped being about Sora
What made Sora worth writing about twice is that it stopped being a story about one app. The cost wall Sora hit first is now the wall every AI company is building against.
The clearest signal is silicon. OpenAI’s $20 billion Cerebras deal was an early move to build non-Nvidia inference infrastructure, and by August, Anthropic had started designing its own inference chip with the explicit goal of cutting per-token cost by roughly half. When two of the three largest labs are pouring capital into custom silicon whose entire job is to make inference cheaper, they are all answering the question Sora asked: how do you serve a compute-heavy workload without losing money on every request?
The other signal is that supply, not demand, has become the ceiling. Nvidia’s own commentary through 2026 has been that memory price increases exceeded expectations and were headed higher, a squeeze I broke down in the AI memory shortage. And the economics that killed a consumer app now shape whether these companies can even go public: OpenAI still loses money per revenue dollar, which is part of why it walked away from a 2026 IPO while Anthropic sprinted toward a $2 trillion listing. Sora was the first public casualty of AI’s cost-per-token problem. It will not be the last decision that problem drives.
What this still changes for anyone buying AI
In 20-plus years running IT operations and, more recently, fractional COO work, the most expensive mistakes I have watched teams make with vendors came from confusing two very different things: a capability a vendor is proud of, and a capability a vendor can actually afford to keep offering. Sora was the second kind, sold as the first. The lesson carries directly into how you evaluate any AI product you are about to build on.
Price in the subsidy ending. If you are getting frontier capability at a price that seems too good, someone is covering the difference, and subsidies get withdrawn the moment the compute is needed elsewhere. Ask which side of your vendor’s income statement your workload sits on. A product line that covers its own inference cost is stable; a loss leader is a strategic decision away from disappearing, and your service level agreement will not save you. Disney had a billion dollars committed and got less than an hour of warning.
Keep your integrations swappable. Abstract the model behind your own interface so you can move providers without rewriting your stack, because the provider most willing to undercut everyone today is the one most likely to reprice or retire the product tomorrow. And watch the direction of investment, not the marketing. The compute rendering hobbyist clips in early 2026 is now running the enterprise APIs and coding agents that businesses actually pay for. Being on the growing side of a vendor’s roadmap matters more than any single benchmark.
The verdict, five months on: OpenAI made the right call on Sora, for the reason it said, and the correction to the famous $15 million figure only reinforces it. What I underestimated was how quickly the same arithmetic would come for everyone else. The free-spending consumer era of AI ended with Sora. The disciplined era it started is now squeezing the enterprise products that were supposed to be immune.