The AI video model is not the product any more

Three things happened this week that look like separate stories about separate companies, and they are the same story.

On September 24, Runway published a benchmark of its own Model Router. Running 250 image-to-video prompts four ways, it found that capping generation at $1.00 cut the cost of a clip by 66%, from $1.80 to $0.61, while holding a 74% usable rate against a Seedance 2.5 baseline of 78%. Then it said something more revealing than the table: 64% of builders using the router on Runway Dev have configured it for cost optimization. That is a customer base telling a vendor, in the vendor's own metric, that it wants the cheap routing path as the default.

On September 23, Comfy launched Comfy Router, an API where the provider is a parameter. The same model — Seedance 2.5, MiniMax H3, Kling, Nano Banana Pro, GPT Image 2, Black Forest Labs on day one — can be called on fal, Runware, WaveSpeed or Higgsfield by changing one line. Comfy's own selling line is that the model string and the provider string are the only two things that change, and that every job reports which provider ran it.

And on September 24, Higgsfield put $20 million behind 100% API cashback on every model in its catalogue, capped at $100,000 per business, with unused credits expiring September 30. Companies do not refund a year's API spend to celebrate a model. They do it to buy volume on a platform where the model is someone else's.

Put together, the picture is not that models stopped improving. It is that model quality stopped being the variable that decides a studio's week. Runway's data says routing rules now move a bill by two thirds while moving usable output by four points. Comfy's product says the choice of host is a string literal. Higgsfield's cashback says the host is willing to pay you to be your default.

If that is right, then a few habits are worth rethinking.

The first is model loyalty. If a routing layer can hold 95% of the usable rate of the state-of-the-art model while most of a prompt set is served by cheaper models, then your aesthetic is being expressed through a rule, not a model. Build the rule — a cost cap, a per-category preference, a fallback — and you stop re-evaluating your stack every launch week.

The second is where your asset value sits. When a clip's cost is a routing decision, the durable things are the inputs: the character sheets, the reference images, the prompt revisions, the keep-and-kill list, the shot plan. Those transfer between hosts. A favourite endpoint does not, and this year alone has shown how fast an endpoint can disappear.

The third is what to measure. "Is this model better" is now a much weaker question than "what does this model cost per usable clip on my prompts". Runway's report is a vendor grading itself; nobody has reproduced it, and its dollars are Runway Dev's prices. But the method is free to copy, and it is the only number that maps to a budget.

None of this makes the frontier irrelevant. Dialogue, lip-sync, thirty-second single takes and physics are real capabilities, and the categories in Runway's appendix show exactly where a cheap route falls apart — 16% usable on dialogue against 92% for the baseline. That is the argument for routing rather than against quality: know which of your shots are the ones that need the expensive model, and let a rule handle the rest.

The uncomfortable part is what it means for the labs. If the routing layer decides what runs, the model becomes a cost of goods that competes on price per usable second, and companies with a hosting business and a cashback budget will keep pushing that price down.

Next test for this argument: watch whether the next round of model launches leads with capability benchmarks or with per-usable-clip costs on somebody else's router.