Models
Every video model CastTake runs, what each costs per second in credits, what a 5-second clip comes to, and which to pick for testing versus promotion.
Last updated
What this page is
Every model this product can run, what it costs you, and where the number came from. A credit costs about 4 cents on the Creator plan and 7 to 10 cents in a credit pack. A generation is quoted before you submit and the credits reserved are that same number, so nothing here is an estimate that gets revised after the fact.
A model that is not on this page cannot be generated. That is enforced in code rather than by convention: the price list is the register of what exists, so a model with no price is unroutable rather than free.
Available models
| Model | Resolution | Credits per second | A 5-second clip | Audio |
|---|---|---|---|---|
| Grok Imagine 1.5 | 480p | 4 | 20 credits | Yes, same pass |
| Grok Imagine 1.5 | 720p | 8 | 40 credits | Yes, same pass |
| Seedance 2.0 Mini | 480p | 3 | 15 credits | Not verified |
| Seedance 2.0 Mini | 720p | 5 | 25 credits | Not verified |
| MiniMax H3 | 720p | 2 | 10 credits | Yes, same pass |
| MiniMax H3 | 1080p | 3 | 15 credits | Yes, same pass |
| MiniMax H3 with reference photos | 720p | 8 | 40 credits | Yes, same pass |
| MiniMax H3 with reference photos | 1080p | 12 | 60 credits | Yes, same pass |
| Kling 3.0 Turbo | 720p | 15 | 75 credits | Not verified |
| Kling 3.0 Turbo | 1080p | 18 | 90 credits | Not verified |
| Veo 3.1 Fast | 720p | 9 | 45 credits | Yes; sold as 8-second clips only, 72 credits |
| Veo 3.1 Fast | 1080p | 9.8 | 49 credits | Yes; sold as 8-second clips only, 78 credits |
| Seedance 2.5 | 720p | 65 | 325 credits | Not verified |
MiniMax H3 at 720p is the cheapest finished five-second clip on this page.
Seedance 2.5 costs more per second than either of the above. It is built for multi-shot scenes with dialogue in a single render and sells at 720p only so far; for single shots, MiniMax H3 costs a fraction as much.
Image models
Still images are billed per picture rather than per second, out of the same credit balance, and quoted the same way before you generate.
| Model | Resolution | Credits per image |
|---|---|---|
| Nano Banana | 1k | 3 |
| Nano Banana 2.1 | 1k | 3 |
| Seedream 5 | 2k | 3 |
| GPT Image 2.5 | 1k | 4 |
| Seedream 4.5 | 2k | 4 |
| Nano Banana 2.1 | 2k | 4 |
| Nano Banana 2 | 1k | 5 |
| GPT Image 2.5 | 2k | 6 |
| Nano Banana 2 | 2k | 7 |
| GPT Image 2.5 | 4k | 9 |
| Nano Banana Pro | 1k | 10 |
| Nano Banana Pro | 2k | 10 |
| Nano Banana Pro | 4k | 13 |
These are what the product-scene tool spends. Finding a reference image is an input to an ad rather than the ad itself, which is why the cheap end of this table matters more than the top: five tries on Nano Banana cost 15 credits against the 40 a 720p clip costs. We do not offer image models that are known to produce explicit images, even when they are cheaper.
Which one to pick
MiniMax H3 for most clips. At 10 credits for five seconds at 720p it is the cheapest video here, so it is where to try ten hooks to find the one that works. It also renders 1080p, makes the sound in the same pass, and accepts a reference image, which is how a presenter stays the same person across a campaign rather than drifting between batches.
Grok Imagine 1.5 when speed matters most. It is the fastest to come back, and at 480p it is already past the resolution most of a phone feed is watched at.
Seedance 2.5 when one clip has to hold several shots or a two-person exchange. It is the priciest model here, so find the idea on H3 first.
Two of our video effects take two reference photos, one per person, and deliver the full video at 720p: the hotel lobby AI video puts two faces at one mic in an orange booth, and the AI APT MV casts one person on drums and the other at the mic. Both pages show full renders and what drifted in them.
The order matters more than the choice: test wide on the cheap model, finish the winner on the expensive one. Rendering every draft at 1080p costs half as much again on H3 and does not make a weak script work.
Where these numbers come from
Every price on this page is the same number the studio quotes on the Generate button, read from one rate card, so this table cannot drift from what you are charged. A model is priced per second of finished video at each resolution it renders, and the price is checked against real generations rather than a published list.
Each resolution is priced on its own rather than by a general rule, which is why the steps between 720p and 1080p differ from model to model.
What is not here
No ranking of the image models against each other. They differ in price and in maximum resolution, both of which are in the table above; which one renders your product better depends on the product.
(Until 2026-09-11 this section said "no image models" outright. That was written when it was true and left standing after the image shelf shipped -- seventeen GPT Image 2.5 generations had been billed in production by the time anyone noticed the page denying they existed.)
No per-model quality claims beyond the table. Which model produces a better ad for your product is decided by the script and the angle far more than by the renderer, and any ranking we published would be marketing rather than a measurement you could check.
AI face swap video: star in a ready-made scenePick a scene, from a product demo to a red carpet, add one photo, and you are the person in it. For creators, shops and anyone with a birthday to celebrate. One payment, no subscription.Try the face swap video tool