AI Video Generation: The Complete Guide

How Video Pricing Works

Before the model table, one rule that governs everything below it: video is priced per second of output on almost every model. A 10-second clip costs roughly double a 5-second clip, and where a model offers several resolutions the higher ones cost more per second. The maths is deliberately simple — Kling O3 Pro is 18 credits per second, so a 5-second clip is 90 credits and a 15-second clip is 270.

VEO 3.1 and VEO 3.1 Fast are the only exceptions: they bill per clip. VEO 3.1 is 225 credits without audio or 338 with, and VEO 3.1 Fast is 120 or 190, charged as a flat rate for a 5 to 8-second generation. Because the price does not move with duration, you should always request the full 8 seconds on those two models. VEO 3.1 Lite is not an exception — it is 10 credits per second, roughly 80 for an 8-second clip.

The exact total is always displayed on the generation page before you commit, so you never have to do this arithmetic under pressure.

Choosing the Right Video Model

Twenty-four models are available on the video generator, each with distinct strengths. The table below summarises the key differences; the prose that follows explains which to actually pick. Credit figures are per second of output unless the row says per clip.

Start with Kling O3 Pro. At 18 credits per second it gives you 1080p, native audio, reference-to-video and first-last-frame across 3 to 15 seconds — a 5-second clip for 90 credits. Nothing else on the platform combines that resolution, that feature set and that price, which is why it is the best overall value pick. Kling O3 at 14 credits per second is the same idea slightly cheaper and is a great default for animating stills, while Kling O3 4K at 50 credits per second is there when you genuinely need 4K.

Gemini Omni Flash is the other value pick, at 15 credits per second for 3 to 10 seconds in 16:9 or 9:16 — 150 credits for a full 10-second clip with audio you direct through the prompt. It is a featured model for good reason: it is the cheapest way to get Google-quality video with sound.

Sora 2 from OpenAI is 12 credits per second, so 4 seconds is 48 credits, 8 seconds is 96 and the 12-second maximum is 144. Its strength is physical realism, which makes it a good choice for shots involving impacts, liquids or objects interacting. Sora 2 Pro is the production tier at 35 credits per second standard, 60 at 1024p and 85 at 1080p — expensive enough that you should only run final selects through it.

The VEO 3.1 family from Google DeepMind is the premium naturalistic option, and the one place where per-clip pricing applies. VEO 3.1 is 225 credits without audio and 338 with; VEO 3.1 Fast is 120 or 190 for close to the same quality. If you have been avoiding VEO on cost, VEO 3.1 Lite changes the picture entirely — 10 credits per second at 1080p with audio, about 80 credits for a full 8-second clip.

The Seedance family from ByteDance spans the widest range on the platform. Seedance 2.5 reaches 30 seconds in a single generation at 60 credits per second at 720p or 30 at 480p. Seedance 2.0 is 35 credits per second at 720p with 21:9 ultrawide, 2.0 Fast is 30 with quicker turnaround, and 2.0 mini is 15 with the same modes. At the value end, Seedance 1.5 Pro is 7 credits per second with native audio and Seedance 1.5 Lite is 5, reaching up to 1080p — a 10-second Lite clip is 50 credits, which is what makes it the drafting model of choice.

The Hailuo models from MiniMax cover three price points: Hailuo 03 at 35 credits per second for up to 2K and 15 seconds, Hailuo Pro at 12 for 1080p, and Hailuo Lite at 8 for 768p with a 1 to 2 minute turnaround. Pro and Lite have fixed 6 or 10-second durations, so the totals are simply 72 or 120 credits and 48 or 80.

WAN 2.6 from Alibaba Cloud is 12 credits per second at 720p or 18 at 1080p with fixed 5, 10 or 15-second durations, multi-shot support and audio — 180 credits for a 15-second 720p clip is exceptional value for that length, and it is one of the strongest models for motion-heavy scenes. Happy Horse, also from Alibaba, is 32 credits per second at 1080p and specialises in expressive, character-driven motion. LTX 2 from Lightricks is 6 credits per second at 1080p, 10 at 1440p and 18 at 4K across 6 to 20 seconds — both the cheapest long clips and the cheapest 4K on the platform.

Camera Movement Vocabulary

Including camera movement descriptions in your prompts dramatically improves video quality. Models have been trained on footage labeled with standard cinematography terms, so using the right vocabulary is one of the simplest ways to level up your results.

Basic Movements

Static shot — the camera does not move; good for dialogue or product showcases. Pan — horizontal rotation, useful for revealing a wide scene. Tilt — vertical rotation, great for revealing tall subjects like buildings or waterfalls. Zoom in / out — changes the focal length to draw attention toward or away from a subject. Push in — physically moves the camera closer, creating a sense of intimacy. Pull back — moves the camera away, often used for dramatic reveals.

Advanced Techniques

Dolly shot — smooth forward or backward movement on a track. Tracking shot — follows a moving subject laterally, keeping pace with the action. Crane shot — sweeps upward or downward from a high angle. Handheld — produces natural, slightly shaky movement for a documentary or found-footage feel. Steadicam — smooth, floating motion often used for walking-and-talking scenes. Drone shot — aerial perspective, perfect for landscapes and establishing shots.

Crafting Effective Video Prompts

A well-structured prompt is the difference between a generic clip and a cinematic one. Think of it as five layers stacked together:

Basic Example

"A person walking down a busy street, natural lighting, realistic." — This works, but the model has a lot of room to improvise. You will get a usable clip, but the framing and mood will be unpredictable.

Advanced Example

"Slow push-in shot of a barista crafting latte art in a cozy coffee shop, warm golden hour lighting streaming through large windows, steam rising from the cup, cinematic depth of field, 24fps, professional color grading." — Every layer of the prompt framework is present: camera movement (slow push-in), subject (barista + latte art), environment (coffee shop + windows), mood (warm golden hour), and technical details (depth of field, 24fps, color grading). The result is dramatically more controlled and visually rich.

Draft on a Cheap Model First

The most valuable habit in AI video is separating the prompt problem from the quality problem. Prompt structure, pacing and composition all read perfectly well on an inexpensive model, so iterate on Seedance 1.5 Lite at 5 credits per second — a 5-second test is 25 credits — until the shot is right, then regenerate the keeper on whichever premium model suits it. The same 5 seconds on Sora 2 Pro at 1080p is 425 credits, so getting the prompt wrong there is seventeen times more expensive.

Once you have your clips, stitch them together in the Video Editor, which is free to use and does not cost credits to export. If you would rather have the structure planned for you, Storyboard and Scene Builder both build multi-shot sequences, and The Lab’s Marketing Room and Video Agent handle guided end-to-end briefs.

Frequently Asked Questions