An AI clip that always finds a way to be made

On a paid plan, a PageToVid film can include AI video clips: short generated shots for the moments a website cannot show, placed between the real recordings. A scene can name a specific model, but most people do not, and they should not have to.

A scene that names no model gets the house clip: our choice of model, at one known price. Since October, that choice had been Veo 3.1 Fast, from Google. It was a good one, and the house clip was built around it.

But a promise to the person making the film should not rest on one provider's catalogue. Video models are young. They are renamed, replaced and retired, often with little notice.

On November 3, a test film came back with a text card where each of its two clips should have been. Our safety net had done its part: a visual that cannot be made never fails a film. The scene shows a designed card instead, the render finishes, and the clip is refunded automatically.

The reason was simple. Google had withdrawn its Veo models from the Gemini API, and our house clip had nowhere left to go. A refund protects the money, but what you came for is the clip. We wanted the house clip to keep its promise whichever model was available on the day.

We already worked with two other video models that answer well. The house clip now treats them as a chain, tried in order:

1. 1Veo, as before, whenever it is available.
2. 2If that gives nothing, Seedance 2.0 Mini: an 8-second clip in 720p, with sound.
3. 3If that fails too, MiniMax H3.

The price of a house clip stays the same whichever route makes it. The order is a setting rather than code, so we can move a provider up or down the list as models change, and switch the chain off entirely if we ever need to.

Our image and video layer also remembers a model that has been withdrawn. After the first refusal, it stops asking that model again, so a long film with several clips does not wait on the same closed door twice.

A fallback can quietly change what someone asked for, so we wrote down what it may and may not do:

• A scene that names a model is never rerouted. A model is a look; if you asked for one, you get it, or the clip is refunded, never a lookalike.
• Every scene records which route made its clip and why, and the film's list of models used names that route.
• The scene itself still says “house clip”, so a later regeneration takes whichever route is healthy at that moment.

Assistants see the same thing. In our MCP server, list_models shows the house clip as available as long as any route can make it, and says which route it will take. Claude or ChatGPT can check that before a film spends anything.

We used to think of the house clip as a model. It is really a promise about a result: a short generated shot that fits its scene, at a price you know in advance. Several providers can keep that promise, and the film should not care which one did.

For you, that means something simple: you choose AI clips, and you get AI clips. Whether Veo, Seedance or MiniMax made them on the day is written in your film's list of models, if you are curious; the film itself just carries on.

The same idea now shapes how we look at every outside service we rely on. A fallback that guards a single call protects against a bad minute; a chain of equivalent routes protects the result itself. You can see what AI clips bring to a film on /features/ai-clips.