What an API dependency really costs
Build your production on a third-party AI API and every generation is a metered call you can't control: the price is theirs to raise, the rate limit is theirs to set, and the model is theirs to retire. By running open-weight models on your own compute, the production layer carries none of that — no per-token bill, no vendor who can reprice or revoke the thing your business depends on.
The production stack, in the open
Open-weight generation
Video, voice, score and lip-sync run on open models — the kind you can inspect and run yourself — deployed onto your own compute.
No per-token meter
Because the models run on your GPUs, there's no per-generation API charge. You pay compute, at cost, once.
Nothing to deprecate
An open weight you hold can't be sunset out from under you. The model you launched on is the model you keep.
Your throughput, your rules
No external rate limit caps your business. Your capacity is bounded by your own compute, which you control.
Inspectable, not a black box
Open weights mean the production stack is knowable — not an opaque endpoint you can only hope keeps working.
Where the line actually is
Straight answer: the platform's intelligence and onboarding layer does use a hosted model to plan and direct work — that part runs on the platform's side, not as a per-generation tax on you. What's free of third-party AI APIs is your production: the generation and rendering that produce your actual output run on open weights on your own compute. That's the part that would otherwise meter you forever, and it doesn't.
Open weights vs. metered API
| Per-generation cost | Compute, at cost | A charge on every call |
| Model lifespan | Yours to keep | Deprecated at their will |
| Throughput cap | Your own compute | Their rate limit |
| Transparency | Open, inspectable | Opaque endpoint |
| Who runs it | Your GPUs | Their servers |
Questions about ownership
So you really use no APIs at all?
Not quite, and we won't claim that. Your production — generation, voice, score, render — runs on open weights on your own compute, with no third-party AI API. The platform's planning and onboarding layer does use a hosted model on our side; it doesn't meter your generations.
Which models actually run on my compute?
Open-weight models for video, voice synthesis, scoring and lip-sync, deployed onto your own Modal and run on your GPUs.
Why is avoiding per-token billing such a big deal?
Because it's the cost that never stops. A metered API charges you on every generation forever; open weights on your own compute turn that into one compute bill you control.
Can the models I launched on disappear?
No. Open weights you hold can't be deprecated by a vendor. The production stack you start with stays available to you.
Own it, don’t rent it.
Your compute. Your data. Your outputs.
See how the Business Builder works →