What Actually Happened
Two things happened in the AI video market in the span of a few months, and read together they tell you exactly where not to put your money.
First, the most famous product died. OpenAI discontinued the Sora web and app experiences on April 26, 2026, and its API is scheduled to shut down on September 24, 2026, according to OpenAI's own help center. The reason was not a lack of hype. Wall Street Journal reporting cited by Futurum put Sora's running cost at roughly a million dollars a day against about $2.1 million in total in-app revenue over the product's entire life. The best-known name in generative video could not make the math work, so it is being switched off.
Second, the challengers are flush. On July 5, 2026, Kling, the video model from Chinese lab Kuaishou, closed a raise of about three billion dollars, the largest single round for a video-generation model this year and second only to DeepSeek across China's generative-AI sector, per Seoul Economic Daily. That capital is aimed squarely at ByteDance's Seedance, which took the top of the text-to-video leaderboard earlier in the year.
And the leaderboard itself has tilted. As of mid-2026, seven of the top eight text-to-video models are Chinese, with Kling and Seedance trading the number-one spot depending on which arena you read (Artificial Analysis). The Western tool most B2B marketing teams recognized by name is the one being turned off.
Put those three facts next to each other. The famous model is dying, the leaderboard reshuffles by the month, and the biggest checks are going to labs that will trigger a very different procurement and data-governance conversation inside a US enterprise. That is not a market you standardize on. That is a market you rent from, carefully.
Why The Model Is The Wrong Thing To Standardize On
Standardizing your production pipeline on one generative model is a continuity risk dressed up as a technology decision. It feels like diligence and behaves like exposure.
Marketing teams love to pick a winner. It feels rigorous to run a bake-off, choose the model with the best benchmark, and build your templates, prompts, and workflows around it. But the last two quarters just proved the winner is not stable. A model with OpenAI's brand, budget, and distribution got discontinued. A challenger raised three billion dollars almost overnight. The thing you would have standardized on in January is not the thing leading the arena in July.
There is a second problem that stays invisible until legal or security asks about it. When the top of the leaderboard is dominated by models trained and hosted by Chinese labs, a company selling into regulated industries, government, or the enterprise inherits a data-governance question every time that model touches a client asset. Where does the footage go, who can see the prompt, whose servers render the frame. Those questions do not have to be dealbreakers, but they turn a casual tool choice into a procurement review, and a procurement review is not something you want load-bearing in your weekly content calendar.
So the model is the wrong layer to make permanent. It is the fastest-moving, most fungible, most politically exposed part of the entire production stack. What should be permanent is everything the model never touches: the footage you shot, the customer who went on camera, the transcript, the brand system, the editorial judgment about which forty seconds carry the argument. Those do not get discontinued. Those do not raise a round and pivot. Those are yours.
The uncomfortable implication is that model-shopping, the activity that feels like staying current, is actually the low-value work. Swapping Veo for Kling for Seedance is a Tuesday-afternoon config change. The work that compounds is the work that survives the swap.
The Data We Actually See
Start with our own book, and treat it as operator observation rather than a study, because that is what it is. We are a studio, not a research lab.
Across our retainer work with 20-plus B2B SaaS engagements, we have changed the primary generative model in our pipeline three times in the last eighteen months. Not one client noticed, and not one deliverable had to be rebuilt, because the model was never the thing the client owned. It rendered a b-roll insert or a background plate, and when a better or cheaper option appeared we swapped it the way you swap a stock-music provider.
The second number is about proportion, and it is what keeps us calm when the leaderboard churns. Across our library, we have never shipped a client engagement where model-generated footage exceeded roughly a fifth of finished runtime. The other four-fifths is real: a founder explaining a tradeoff, a customer describing the exact moment something worked, a recorded demo, a room at an event. That ratio is deliberate. It is the design choice that makes the model a garnish instead of the meal.
Now set that against the public record. Sora cost about a million dollars a day to run and returned $2.1 million across its entire life before OpenAI pulled it (Futurum). Kling just raised roughly three billion to keep competing (Seoul Economic Daily). Seven of the top eight models are Chinese, and the rankings move month to month (Artificial Analysis). Every one of those facts argues for keeping the model swappable and the ownership somewhere else. The teams that hard-wired their whole workflow to Sora's API are now running a migration on OpenAI's timeline. The teams that treated it as one rendering option are changing a setting.
The Strongest Case Against This
Here is the honest counter, because there is a real one. Standardizing on a single model is not irrational. It buys you a consistent look, reusable prompts and templates, a team that gets fast because it is not relearning a tool every quarter, and sometimes volume pricing. Constantly swapping models has a cost, and pretending it is free would be dishonest.
That is true, and it is why the argument is not that you should never commit to a model. The argument is narrower: commit at the workflow layer, not the vendor layer. You can absolutely standardize on how you brief, how you storyboard, how you review, and how you fold generated footage into a real edit. Those habits transfer across models with a little friction. What you should not do is let one vendor's API become the thing your entire output depends on, because that is the single dependency the last two quarters proved you cannot control.
There is also a quality objection worth taking seriously. Some of these models are genuinely excellent, and Kling's motion or Seedance's audio can do things that were impossible a year ago. Fair. Use them. Use the best one available this month. The point is not to avoid good models. The point is to enjoy them without marrying them, so that when the leaderboard reshuffles again, which it will, you are upgrading rather than rebuilding.
What To Do Monday
First, audit your dependency. Write down, honestly, what breaks if your current generative model disappears in ninety days the way Sora's API is about to. If the answer is a lot, you have standardized on the wrong layer.
Second, keep the model swappable. Treat generative video as a rendering option that sits behind your workflow, not as the workflow itself. Anything reusable, your brief format, your prompt library, your review process, should live above the model so you can change the model without changing the system.
Third, protect the ratio. Make sure the majority of every deliverable is footage you own and could never regenerate: real people, real demos, real rooms. The generated frames are the seasoning, not the substance, and that ratio is what turns a model swap into a non-event.
Fourth, get ahead of the governance question before procurement does. If you sell into regulated or enterprise buyers, decide now where client footage and prompts are allowed to go, and choose models accordingly. It is cheaper to answer that question on your own schedule than in a security review in the middle of a deal.
Do those four things and the next Sora shutdown or the next three-billion-dollar raise becomes a headline you read with interest, not an emergency you manage. The model was always the rental. The footage and the system were always the point.