MiniMax H3 Max Recast on fal: Price, Limits and Whether It Can Replace the Person in Your Reel (October 2026)
Recast swaps people in an existing clip using one to four reference photos and keeps the cuts and the soundtrack. What it costs, what it accepts, what it does with several people in frame, and what we have and have not tested.
MiniMax H3 Max Recast is a video-to-video model for one job: take a clip that already exists, replace the people in it with people you supply as photos, and leave everything else alone. It appeared on fal at the start of October 2026, and it is the first model we have seen that lets you say who becomes whom. That makes it the most interesting new entrant for anyone trying to put their own face into someone else's Reel. It also means a lot of claims are circulating that nobody has checked. This guide separates what the schema and the price pages say from what is still unproven.
What it takes and what it returns
The input is a source video between 5 and 30 seconds long, with one rule that matters: no single shot in it may run longer than 15 seconds. You pass one to four reference images (reference_image_urls), one per person you want to substitute. Replacement runs from left to right across the frame, and the prompt can override that order, for example "use the first photo for the presenter on the right". There is an enhance_realism switch, and the default resolution is 1080P with 768P available.
Two behaviours are the reason people care. It keeps the cuts: a Reel with eight edits comes back with eight edits, in the same places, so you do not have to chop the source into scenes and stitch the results. And it keeps the audio of the source, so a trending sound survives the swap. Most replacement models regenerate a clip scene by scene and lose one or both.
What it costs
fal's own pricing data shows a base unit of $0.05, which understates the real bill. Three third-party price trackers report the actual rate per second of output as $0.30 at 768P and $0.45 at 1080P. We could not read fal's model page directly, so treat those as reported rather than confirmed.
| Output length | 768P | 1080P |
|---|---|---|
| 5 seconds | $1.50 | $2.25 |
| 15 seconds | $4.50 | $6.75 |
| 30 seconds (the input maximum) | $9.00 | $13.50 |
For comparison, the video-edit model our autopilot runs today, HappyHorse video-edit, bills input and output seconds together and works out at about $0.28 per delivered second at 720p. At 768P the two are close; at 1080P Recast is roughly 60 percent dearer. The question that decides the economics is not the rate but how many attempts a usable result takes, and that is a number only a test can give.
Where it could beat what we run today
Our current pipeline cuts a Reel into chunks of at most 15 seconds, edits each chunk and joins them. Chunk boundaries are where clothes and hair drift, which we documented in Kling 3.0 vs HappyHorse for character replacement. A model that accepts a 30-second source with its cuts intact removes the chunking, and with it the boundary problem. The explicit person-to-photo mapping also turns our biggest warning, "every face in this Reel will become your avatar", into a choice. We cover that case in who gets replaced when there are several people.
What we have not verified
This is the part to read twice.
- Identity. No one has certified any video model for keeping a face recognisable, and fal's own model recommender returns "inspection required" for all of them. We have not run Recast on our own assets yet. A blind three-way test against HappyHorse and Kling O3 edit is next on our list, with each run approved separately because each costs money.
- Real faces. MiniMax's content filter, error code 1026, rejects named public figures and protected characters. Whether it accepts an ordinary person's photo reliably is unknown to us.
- Motion and lip sync. A source with fast movement or someone speaking to camera is exactly where any replacement model slips. Expect to try short clips first.
- Date and price. The early-October release date and the per-second rates come from third-party sources, not from fal's pages.
Who should try it directly
If you are comfortable with an API and have a short source clip you have the right to use, a five-second test costs about $1.50 at 768P and tells you more than any article. Use a clear frontal photo, keep the person in the source reasonably large in frame, and start with one person.
If you would rather not, the Studio's autopilot does the equivalent job today with HappyHorse video-edit; we will add Recast as an option only if it wins the test, and we will say so here when it does or does not.
Common questions
Is Recast the same as a face swap?
No. A face swap pastes a face onto the existing performer. Recast regenerates the person from your reference photos, so clothes, build and movement follow your references while the scene, the cuts and the sound of the source clip are kept.
Can it replace only one of two people?
That is what the explicit photo-to-person mapping is for. You supply one photo and, if needed, a sentence such as "replace the presenter on the right". We have not yet verified how reliably it obeys that on real footage, so run a five-second test first.
How long can the source clip be?
Between 5 and 30 seconds, with no single shot longer than 15 seconds. A Reel with frequent cuts usually satisfies that; a 30-second single take does not.
Is it available inside AutoHustle?
Not yet as an option. Our autopilot uses HappyHorse video-edit today. Recast is a candidate, and we will only add it if it beats what we run in a blind test on clips we own.
How to do this without code: Replacing the person in a Reel is the job of the autopilot, and Recast is a candidate we are testing for it. In AutoHustle this is the autopilot: paste the Reel link, add one photo of yourself and confirm the plan (about 4 credits per second of finished video, so 60 credits for a full 15-second call; analysing the Reel is free). See How to put your face in any Reel and Recreate an Instagram Reel.
Related reading: MiniMax H3 Max Image-to-Video on fal: Price, Limits and Examples (October 2026) · MiniMax H3 Max Lip Sync on fal: Talking Photos From Audio Up to 15 Minutes (October 2026)
For developers: the fal endpoint
| Model | fal endpoint |
|---|---|
| MiniMax H3 Max Recast | minimax/h3-max/recast |
Inputs: video_url (5–30 s, no shot over 15 s), reference_image_urls (1–4), prompt, resolution (768P or 1080P, default 1080P), enhance_realism. fal lists a base unit of $0.05; the per-second rates above are third-party figures. Read the live price with fal's pricing endpoint before you build.