Sora's API Is Shutting Down as Higgsfield Hits $5.4 Billion: The AI Video Model Isn't the Product Anymore

AI video models keep getting better. As products, they keep getting harder to find on their own.

The old way of thinking about this market is easy to love. A lab trains a model, the model makes clips nobody else can make, and creators line up to pay that lab. When ByteDance's Seedance 2.5 launched on July 31, 2026 and doubled single-pass generation from 15 to 30 seconds, the appeal of that story was obvious. One model moved, and every storefront in the business rearranged its shelves.

But the story decides where creators put their subscription money and their muscle memory, so it matters whether it is still true. On September 24, 2026, OpenAI will remove its Videos API and every Sora 2 model from its platform, according to OpenAI's deprecations page. Other retirements on that page list a recommended replacement. The Sora rows show only dashes. The Sora app itself was discontinued on April 26.

The Case for the Model Makers

The strongest argument for the labs is that frontier quality still decides what is possible. No router can add 15 seconds to a clip. A lab had to do that, and an aggregator's menu is only ever as good as what the labs choose to cook.

The labs also own the scarce thing. Training a frontier video model takes compute and data on a scale that storefronts cannot match. ByteDance recently secured a $30 billion loan as it expands its AI capabilities, Bloomberg reported, in a story republished by The Star. Higgsfield's entire Series B is $400 million.

A lab can also build a real business on its own work. Runway reached $200 million in annual recurring revenue this month, double its level five months earlier, and most of that comes from its video-generating models, co-CEO Anastasis Germanidis told Bloomberg (via Investing.com's summary).

Then there is dependency. Aggregators rent what they sell. Pika's pitch for its relaunch leans on offering Seedance at half competitors' prices, Semafor reported, which is a pitch built on ByteDance's model. OpenAI has just shown that a lab can withdraw access on six months' notice; its page says developers were told on March 24.

And nothing stops a lab from building the editor too. Runway now runs as a panel inside Premiere Pro and After Effects.

All of that is true, and the money still went the other way.

Why the Routers & Editors Win Anyway

Start with the numbers. Higgsfield raised $400 million at a $5.4 billion valuation on August 17 and said its annualized revenue had reached $700 million. Higgsfield's homepage, as of September 19, leads with Seedance 2.5 and Google's Nano Banana Pro, both other companies' models, and advertises more than 50 models through one API.

Runway's last reported valuation was $5.3 billion, per TechCrunch, citing a source familiar with the matter. So a company whose front page leads with other labs' models is valued slightly above Runway and reports 3.5 times the run rate ($700 million divided by $200 million). Both figures are the companies' own and are not identical measures, but the gap is not subtle. Pika, which trains its own Pika 2.5 model, is valued at $470 million, according to Semafor. That is less than a tenth of Higgsfield's mark ($470 million divided by $5.4 billion is 8.7%).

Now look at what the labs themselves are doing. Runway's API changelog shows third-party models arriving on September 23, 2025, starting with Google's Veo 3. Since August 5 of this year, the changelog adds MiniMax's Hailuo 3.0, Seedance 2.5, Grok Imagine Video 1.5, Alibaba's WAN 3.0 and MiniMax H3 Max. On July 23, Runway shipped a Model Router that chooses by cost, latency or quality, in the changelog's words, "so you never have to name a model."

When a lab's own documentation tells developers the model's name no longer matters, the argument is mostly over. Runway's Adobe plugin makes the same concession: it lists Seedance 2.5, Kling 3.0 Pro and Veo 3.1 beside Runway's Gen-4.5.

Pika made the same move on September 17. Its rebuilt product automates which model generates each element of a video, Semafor reported, and Pika's homepage, as of September 19, lists Seedance 2.5, MiniMax H3, Wan 3.0 and a Google model next to its own Pika 2.5.

Adobe may be the clearest case. Adobe has its own Firefly model and still shipped Premiere's new Generative Media Tool on September 8 with Google Veo, Kling, Runway and Luma in the same menu. The company that owns the timeline decided the model should be a dropdown.

None of this makes the steelman wrong about quality. It makes quality a rotating title. The lab in front changes, and the router simply carries whoever leads this month, while a creator subscribed to a single lab has to move every time. Developers who built directly on Sora have until September 24 to rebuild on another vendor. A developer behind a router would, in principle, change a setting.

The scarcity argument fails in a similar way. Compute is scarce, but a scarce input is not a product. Seedance 2.5 is spreading by being everywhere at once: on Runway's API, in Runway's Adobe panel, on Pika and on Higgsfield's front page. That is the behaviour of a supplier. Higgsfield CEO Alex Mashrabov argued in his company's funding release that the value now accrues to the applications that put models to work, and the labs' own shelves suggest they agree.

Creators should choose tools the way this market now sells them: by workflow, editor and price per second, with the model as a setting that will change. Loyalty to one lab's subscription should have to justify itself every month.

The labs will keep deciding what AI video can do. The routers and editors will decide where it gets done, and who gets paid for it.