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.

First, normalize the price

Video models bill in three ways:

• Per video second, set by resolution. Kling, Wan, MiniMax and HappyHorse.
• Per 1M tokens, set by resolution. The Seedance family. Token count depends on the actual request, so you need an estimate per job.
• Per image or per request. Image models such as Seedream 5.0 Lite and Wan 2.7 image.

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.

The spec sheet that matters

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.

Shot type 1: World-building (establishing shots, landscapes, architecture)

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

• MiniMax-H3 at 2K sits between 1080p and 4K in both size and price (about $0.76 per 5s, compared with $1.91 for kling-3.0 at 4K), which suits aerial moves over landscapes. Castaly's landscape example on the homepage runs on MiniMax-H3 at 2K/5s, and so does the community clip "A fleet of hot-air balloons drifts over mist-filled…".
• wan3.0-video-prime at 1080p is what the architecture-and-interiors example uses (a gimbal walk through a sunlit space). It's the most expensive per second here ($1.55 per 5s), so draft on wan3.0-video first. It accepts the same reference types and costs $0.25 per 5s at 480p.
• kling-3.0 at 1080p works for sci-fi establishing shots at a mid-range price ($0.51 per 5s), and goes to 4K when the shot is a hero 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.

Shot type 2: Characters (close-ups, reactions, continuity)

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

• kling-3.0 handles most of Castaly's character-performance examples: the rainy-night-train close-up and a two-scene continuity test with one character description reused word for word (night market, then a rain-streaked window), all at 1080p/5s.
• If you need the same person across many shots, look at reference capacity. kling-3.0-omni and kling-o1 accept up to 7 "elements" or reference images. seedance-2.5 takes up to 50 mixed references and runs up to 30 seconds.
• Continuity tip: anchor the character with odd, specific details ("a small scar above her left eyebrow, a worn olive field jacket"). They hold better than "consistent facial features."

Shot type 3: Action (sports, chases, creatures)

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

• kling-3.0-turbo is used for the sports and action examples (a dunk, a kickflip) at 1080p/5s for about $0.64. It takes a first frame as its only reference, which suits a locked starting pose.
• kling-3.0 handles fantasy action: "a traveler in a burgundy cloak sprints across a stone bridge as a dragon sweeps low behind her," 1080p/5s.
• seedance-2.5 is the pick for long, one-take spectacle: 10-second 1080p continuous takes where a trigger event escalates (pages becoming a paper dragon, a spilled cup becoming a subway waterfall). Write it as a shot list in one paragraph: "One continuous take, no cuts", then the event, then where the camera starts, moves and ends, then "slow motion at the moment of impact."

A workflow that keeps cost predictable

1. Block the shot cheaply. kling-2.6 at 720p (~$0.19 per 5s) or wan3.0-video at 480p (~$0.25 per 5s). You're checking framing and motion, not detail.
2. Lock the prompt and seed your references. Save the best still or frame and reuse it as a first frame or reference.
3. Render the final on the model that fits the shot type, at the resolution you'll deliver.
4. Keep the settings with the output. Castaly saves the model and full settings with every result in your library. Whatever tool you use, make sure you can answer "which run was this?"

Doing it from code

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"
  }'

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.