AI video generation: pick the engine, then write the prompt
Every video model is good at a different shot. The expensive part is finding that out by burning credits on the wrong one. This page is the shortcut: what each engine is for, what a second of it costs, and prompts that already work.
Start free - 1 prompt, no cardTulexAI runs every video engine it offers on one Platform Token meter from $23/month: Runway Gen-4 for directable shots, Kling for motion that holds up, Google Veo 3.1 for photoreal frames with sound baked in. The per-second cost of each is shown before a render starts, so choosing an engine is a budget decision you can actually see.
Every video generation model on one meter
10 models, one Platform Token budget, from $23/month. The priciest costs about 28x the cheapest per 1 second, which is why picking the right one per job matters more than picking a favourite.
| Model | Provider | PT cost | Per |
|---|---|---|---|
| Kling 2.5 Turbo | Kling | 110 PT | 1 second |
| Kling 2.1 Master | Kling | 720 PT | 1 second |
| Kling 2.6 Pro | Kling | 200 PT | 1 second |
| Kling 3.0 | Kling | 290 PT | 1 second |
| Runway Gen-4 | Runway | 40 PT | 1 second |
| Google Veo 3.1 | 1100 PT | 1 second | |
| Sora | OpenAI | 300 PT | 1 second |
| Sora 2 | OpenAI | 300 PT | 1 second |
| Seedance Pro | ByteDance | 200 PT | 1 second |
| Seedance 2.5 | ByteDance | 850 PT | 1 second |
Which model for which job
A cheap draft to check the idea before committing
Runway Gen-4 · 40 PTThe lowest per-second cost of any engine here, so iterating on framing and pacing barely moves your meter. Render the idea rough, then re-render the keeper on something pricier.
Motion that has to stay physically believable
Kling 3.0 · 290 PTKling holds object permanence and weight through camera movement better than the cheaper engines - the failure mode you notice on a hand, a walk cycle, or anything with a hard edge crossing frame.
A photoreal hero shot with sound already in it
Google Veo 3.1 · 1100 PTVeo 3.1 renders dialogue, ambience and effects with the picture, which removes an entire post step. It is also the priciest second here, so use it on the one shot that carries the piece.
A shot you will steer, then hand to an editor
Runway Gen-4 · 40 PTRunway is the only one of these built around controls rather than a prompt box - camera paths and motion brush mean the second attempt is a correction, not a re-roll.
Volume - lots of clips where none is precious
Kling 2.5 Turbo · 110 PTTurbo trades some fidelity for a fraction of the tokens of Kling 3.0, which is the right trade when you need twelve background plates and not one perfect frame.
Prompts that already work
Copy one, change the subject, keep the structure - the camera, light and constraint clauses are what make these produce a usable result.
Product hero, slow orbit
Google Veo 3.1A slow 180-degree orbit around a matte black ceramic coffee cup on a polished concrete surface. Soft overcast key from camera left, single hard rim light from behind. Shallow depth of field, subtle steam rising. Locked exposure, no camera shake. 5 seconds.
Naming the light direction and the lens behaviour is what separates a product shot from a stock-looking render. "Locked exposure, no camera shake" suppresses the drifting auto-exposure most engines add by default.
B-roll that cuts cleanly against real footage
Kling 3.0Handheld mid-shot walking through a covered market at golden hour, shot on 35mm, natural grain, slight gate weave. Foreground stalls pass through frame left. Ambient crowd, no dialogue. Consistent colour temperature throughout.
Generated clips usually betray themselves by being too clean. Asking for grain, gate weave and a consistent colour temperature makes the clip gradeable alongside camera footage instead of standing out in the timeline.
Looping background plate
Kling 2.5 TurboStatic locked-off wide of drifting fog over still water at dawn. No subject, no camera movement, no cuts. Even exposure edge to edge so the first and last frame match. 4 seconds.
Loops break where the first and last frame disagree. Locking the camera and asking for even exposure edge to edge gets you a clip you can actually cross-dissolve into itself - and on the cheapest engine, because a plate does not need fidelity.
Talking-head alternative with no presenter
Runway Gen-4Close-up of hands sketching a wireframe on grid paper with a fineliner, top-down, warm desk lamp. Pen stays in frame, paper does not shift. Quiet room tone.
Hands are where video models fail most visibly, so constrain them: one action, top-down, nothing else moving. This is the shot that replaces a piece to camera when you have no presenter.
Fixing a shot that came out too busy
Kling 2.6 ProSame scene, but remove all background movement. Single subject centred, static camera, plain out-of-focus backdrop. Nothing enters or leaves frame.
The most useful prompt in any library is the corrective one. Video models over-animate by default; subtracting motion explicitly is faster and cheaper than describing the whole scene again from scratch.
How to get a usable clip on the first budget
- 1
Draft on the cheapest engine first
Render your idea on Runway Gen-4 at the shortest duration the shot allows. You are checking composition and pacing, not fidelity, and this is the least expensive second on the plan.
- 2
Fix the shot, not the prompt
If the draft is close, change one thing - the camera move, the light direction, or the subject count. Rewriting the whole prompt changes several variables at once and you lose the thing that was working.
- 3
Re-render the keeper on the right engine
Move to Kling 3.0 when the motion has to hold up, or Google Veo 3.1 when the frame has to be photoreal and carry its own audio. Only the shot you are keeping earns the expensive seconds.
- 4
Watch the meter before you commit
Every model shows its Platform Token cost per second before the render starts, and duration multiplies it linearly. A 10-second Veo shot is not slightly more than a 4-second one - check the number, not the instinct.
Craft notes
- 01Cost scales linearly with duration but quadratically with resolution - doubling the pixel height roughly quadruples the tokens. Draft at the lower resolution, finish at the higher one.
- 02Describe light before subject. "Soft overcast key from camera left" changes the output more than another adjective about the subject.
- 03Say what should NOT move. Video models default to animating everything; "static camera, background does not move" is often the whole fix.
- 04Generated clips look fake because they are too clean. Ask for grain, a little weave, or a consistent colour temperature and they cut against real footage.
- 05Short beats long. Two four-second clips you can cut between usually beat one eight-second take, and cost the same.
Frequently asked questions
It depends on the shot rather than the leaderboard: Google Veo 3.1 for photoreal frames with native audio, Kling 3.0 for believable motion, Runway Gen-4 for shots you need to steer and cheap drafts. TulexAI runs all of them on one meter, so the choice is per clip.
Every video generation model, one bill
10 models on one Platform Token budget from $23/month - with the per-1 second cost shown before you spend it.
Start free1 free prompt · no credit card · cancel any time