The short answer, by job
| If you need | Use | Because |
|---|---|---|
| New shots at the lowest price | Wan 3.0 | $1.00 per 5 s at 1080p, and it followed a long 30-second prompt on its first run |
| To change a location or recast footage | Kling 4.0 Edit | Kept every cut and move, with the strongest location swap |
| Background people in that same edit | Wan 3.0 edit | Recast extras best, in runs of two or three shots |
| A real person to stay recognisable | Seedance 2.5 Video Edit | The only route that kept the face at full size |
| Action and rivalry with a timed direction | Seedance 2.5 | The selected take of six, though only as a 480p draft |
Specs and price
From vendor and fal model pages on 2 November 2026, plus the files our own runs produced. Prices are list prices for 5 seconds of 16:9 video.
| Kling 4.0 | Seedance 2.5 | Wan 3.0 | Veo 3.1 | |
|---|---|---|---|---|
| Price at 720p | ≈ $0.79 | ≈ $1.16 BytePlus, ≈ $2.31 fal | $0.50 | $2.00 with audio, $1.00 without |
| Price at 1080p | ≈ $1.18 | ≈ $2.84 BytePlus, ≈ $5.69 fal | $1.00 | $2.00 with audio, $1.00 without |
| Editing price, 5 s in and 5 s out, 1080p | $1.41 | ≈ $3.40 BytePlus | $2.00 | Not compared |
| Top resolution listed | 1080p | 1080p on the routes we use; 4K claimed by CapCut | 1080p | 4K |
| Frame rate | 24 fps in our file | 24 fps by fal’s formula | 30 fps in our file | 24 fps |
| Longest single generation | 30 s | 30 s | 30 s | 8 s, extendable in 7 s steps on fal |
| Text, image, reference | All three | All three, up to 50 references | All three | All three |
| Sound in the same pass | Yes, always on | Yes, no extra charge | Yes, no extra charge | Yes, at double the price |
| Speech | Yes; custom voice via a created voice id | Yes; dialogue in quotes; voice reference | Yes; voice reference up to 15 s | Yes |
Sources: Kling 4.0, Seedance 2.5, Wan 3.0 and Veo 3.1 on fal, plus Google DeepMind’s Veo page. fal also marks that Veo 3.1 endpoint for deprecation on 17 November 2026, so expect those rows to change.
What happened in our own tests
These are real jobs with one or a few runs each, so read them as cases and not statistics. Follow the links to see the actual takes and prompts.
Character and object consistency
In the 16-second recast, all routes received the same approved pictures. Kling 4.0 Edit gave a good likeness but left one background extra. Wan 3.0 gave the best extras. Seedance 2.5 in Dreamina got the people and places right, but it re-staged the shot, because a video plus pictures in that mode generates a new clip. In the talking-head test, on the other hand, Seedance’s dedicated Video Edit mode was the only route that kept a recognisable face.
Camera movement and physical realism
Seedance 2.5 made the sideways move and facial effort read best in the running commercial, and its 480p draft threw the cup at the rear window where Google Flow threw it at the camera.
It can also slip into slow motion: long negative lists and words like “cinematic” did that to one of our takes.
For cause and effect, we ran a crane lifting a ship: on the same frames and prompt, a paid Seedance run kept the hook hitting the water 2 seconds before the ship rose, and Veo 3.1 Lite in Flow had the ship rise on its own.
Prompt accuracy
Wan 3.0 followed a beat-by-beat prompt closely in the dollhouse scene and in One Prompt Studio, including a rising studio and a walk into it. It starts skipping instructions when overloaded: a run with six shots and nine pictures recast the people and left the old shop in place. Two shots with two pictures came out clean. Kling 4.0’s first 30-second take for One Prompt Studio was rejected, but it ran on an earlier prompt, so this is no fair head-to-head.
Audio and lip sync
In the dollhouse scene Wan 3.0 delivered both spoken lines, and it added a few notes of music we had ruled out. Google Flow’s first take gave both lines to one character. The dialogue guide covers this properly.
Failed generations and retry costs
Reference-mode Seedance 2.5 on Higgsfield failed twice after about 45 seconds and was refunded, then the same job worked in Video Edit mode. fal refused a real person’s clip on Seedance 2.5, at no charge, so we route those elsewhere. The 30-second cost guide covers what retakes do to the bill.
Cost per usable second
The number that matters is price times takes. We have one clean figure: One Prompt Studio’s two takes cost $13.05 for a 23.6-second cut, about $0.55 per usable second. We have not run the four models on the same brief under the same rules, so we will not give a ranking. Treat any site that gives one without showing its takes with care, especially if it sells access to the winner.
Where other models fit
MiniMax H3 full held up against Wan 3.0 in the dollhouse scene, on our own Hugging Face copy. Turbo is a draft model. Runway Aleph 2.0 and Luma Ray 3.2 appear in the object-editing and reframing notes, and the roundup lists the rest, including why Sora is no longer an option.
Choose the model per shot
The AI Video Production Lab works through references, camera moves and failed takes on your own shots, across these models.
Find your workshop ↗