How to choose an AI video model for each shot type: a spec & price comparison

Most "which AI video model is best?" threads end with "it depends." Here is what it actually depends on. A short film needs different shots: a wide establishing shot that sells a world, a close-up where a face has to stay the same face, and an action beat where bodies and physics have to hold together. Each of those puts different demands on a model, and the models are priced in units that don't compare directly (per second, per million tokens, per image).

This post is the comparison I wanted when I started: specs and public prices for the video models I use, grouped by shot type, plus a simple draft-then-final workflow that keeps costs predictable. I'm using Castaly (a web studio I built), which lists these models side by side with their prices, but the reasoning applies wherever you run them.

To compare per-second models, multiply by your clip length. Here is a 5-second clip at public prices:

The underlying rates, for reference: kling-3.0 is $0.07644/s at 720p, $0.10192/s at 1080p and $0.3822/s at 4K. With native audio it's $0.11466/s at 720p and $0.15288/s at 1080p. wan3.0-video is $0.0494/s at 480p, and MiniMax-H3 is $0.1521/s at 2K.

Seedance is token-priced: seedance-2.5 is $13.2145 per 1M tokens at 480p/720p and $14.4495 at 1080p. seedance-2.0 is $8.645 (480p/720p) and $9.5095 (1080p). seedance-2.0-fast is $3.4357 and seedance-2.0-mini is $1.7086 per 1M tokens at 720p. Because the token count depends on the request, Castaly shows an estimate for the spec you selected before you generate, and the Creations page lists estimate and actual cost side by side after each task. Check that number before you queue a batch of long Seedance clips.

Price is half the decision. The other half is what the model accepts and produces:

Two things stand out. Clip length ranges from a 10-second cap (Kling 2.x) to 30 seconds (Seedance 2.5, Wan 3.0). Reference capacity ranges from one first frame to 50 mixed references. Those two numbers usually matter more than a model's general reputation.

What you need: scale, atmosphere, a slow and readable camera move, and often a larger frame.

Prompt tip: describe the camera path explicitly ("slow push-in past the astronaut toward a monumental ring structure") and keep one dominant light source.

What you need: stable facial features across shots, subtle performance, and natural blinks and breathing.

What you need: fast motion that stays coherent, believable physics, and slow motion when it matters.

Everything above can be scripted. The same models are available through the Modelflare API gateway. Video is asynchronous: submit, then poll.

curl https://modelflare.dev/v1/videos \ -H "Authorization: Bearer $MODELFLARE_API_KEY " \ -H "Content-Type: application/json" \ -d '{ "model": "seedance-2.5", "prompt": "A traveler runs across a stone bridge as a dragon sweeps behind her", "duration": 5, "resolution": "1080p" }' Enter fullscreen mode Exit fullscreen mode Studio and API share the same public prices and the same USD wallet. It's pay-per-use: no subscription, no minimum, and no free tier. If you'd like to see the outputs referenced in this post with their full prompts, they're in Castaly's community gallery, which you can browse without signing up.

👉 Compare the models and prices in Castaly

Prices are public list prices as of October 20, 2026, and change over time. The pricing page is the source of truth.