Anyone producing AI video quickly learns that cost is not in the approved clip but in the discarded ones. In a multi-scene piece it is common to generate three or four versions of each scene until framing, motion and light are right. If every attempt runs on the most expensive model, the bill explodes. The way studios and agencies handle it is the model ladder: explore on the cheap rung, finish on the expensive one.
The rungs, with measured costs
Values from the Oct 10, 2026 GenBrasil test, for a 5-second clip (4 s on Veo):
| Rung | Model | Credits | Time | Use |
|---|---|---|---|---|
| 1. Sketch | Veo 3.1 Lite | 13 | 45 s | framing, idea, pace |
| 2. Draft | Seedance 2.0 Mini | 25 | 1 min 26 s | scene motion and character |
| 3. Pre-final | Seedance 2.0 Fast | 30 | 2 min 37 s | client approval |
| 4. Final | Seedance 2.5 | 76 | 3 min 38 s | delivery |
The math for a 6-scene piece
Illustrative assumption: 6 scenes of 5 seconds and, on average, 3 attempts per scene until it works.
| Strategy | Math | Credits |
|---|---|---|
| Everything on the final model | 6 scenes × 3 attempts × 76 | 1,368 |
| Ladder (sketch on Veo Lite, final on Seedance 2.5) | 6 × 3 × 13 + 6 × 76 | 690 |
The ladder spends about half. At the Pro plan's R$ 0.25 per credit (R$ 299.99 for 1,200 credits), that is roughly R$ 172 against R$ 342 for the same piece. Savings grow with the number of attempts: the more you redo, the more the ladder pays.
Two notes
Seedance 2.5 is included in the Pro and Ultra plans; on Starter and Free, only with extra credits. And the Free plan gives access to light models such as Veo 3.1 Lite and Seedance 2.0 Mini, which already lets you do the whole sketch phase at no cost.
How to apply it without losing style
- Freeze the prompt. Use the same text on every rung, changing only the model. That way the draft predicts the final.
- Freeze the starting point. Use the same image as first frame on every rung. See how to produce it in character LoRA on GPU.
- Approve scene by scene. Don't generate the whole piece at once: one bad shot should not cost the whole piece.
- Only move up with an approved scene. If the motion is wrong in the sketch, it will be wrong in the more expensive final.
What the ladder doesn't predict
The final model has its own style: texture, light and motion may come out different from the draft, and a detail that worked in the light model may fail in the heavy one. Reserve one or two extra attempts on the final rung. Each rung's cost can change, and the current value appears before generating.
Variation: same model, lower resolution
Another ladder is resolution. In a GenBrasil measurement, the same prompt on Seedance 2.5 cost 34 credits at 480p and 76 at 720p. Testing at 480p and finishing at 720p keeps the model's “personality”, with the advantage of not changing style between rungs.
To see current values and start with the sketch, open GenBrasil (Portuguese interface): the credit cost is shown before each generation.
Frequently asked questions
Does the model ladder really save money?
In the example (6 scenes, 3 attempts per scene) it spends about half the credits. Savings depend on how many attempts you make: the more you redo, the more you save.
Will the final look the same as the draft?
Not necessarily. Each model has its own style. Use the draft to approve framing and motion, not the final texture.
Can I do the sketch phase for free?
Yes. The Free plan includes light models and lets you explore ideas without a card, with a watermark and no commercial use.
Make your video on GenBrasil
Eight AI video models in one account, with the cost shown before you generate and payment in reais by Pix or card. The interface is in Portuguese. The Free plan needs no card.
Go to GenBrasil ↗Rather build your own workflow? See the hourly GPU templates.
Read next: the eight models in a real test · how to build your own benchmark