The Token Math Behind AI Video Models: What Changed In Late 2026

A year ago, an AI video model was judged on one question: does the clip look real? The releases that landed this autumn, ByteDance’s Seedance 2.5, MiniMax’s H3 family and Alibaba’s Wan 3.0, have moved the argument somewhere more technical. They all render their own soundtrack. Two of them hold one continuous take for thirty seconds. They accept trays of reference images, clips and audio. And they are priced in ways that only make sense once you see the arithmetic underneath.

I build products on top of these APIs, which means I read their price pages and logs more closely than is healthy. Here is what the numbers say about where video generation is heading.

Video is billed in tokens, and tokens are pixels

The cleanest way to understand AI video pricing is the formula ByteDance publishes for Seedance:

tokens = width × height × frames per second × duration ÷ 1024

Each 1024 pixels of each frame counts as one token. Run it for one second of 24 fps video:

On fal, Seedance 1 Pro Fast is sold at $1 per million video tokens. That works out to about $0.0097 a second at 480p, $0.0216 at 720p and $0.0486 at 1080p, which are exactly the per-second prices fal lists. Seedance 2.5 on ByteDance’s own Ark platform follows the same curve at a higher rate: CNY 70 per million tokens up to 720p and CNY 77 at 1080p. That comes to about CNY 0.67, 1.51 and 3.74 a second, or roughly $0.095, $0.21 and $0.53.

The practical consequence is simple. 1080p costs about five times what 480p does, because it has five times the pixels. Clip length scales linearly on top of that. A 30-second 1080p Seedance 2.5 take lists at around $16 before anyone adds a margin. That is why every serious workflow now drafts at 480p and renders the final size only once the shot is approved.

“1080p” no longer means one thing

The headline resolution on a model card hides three very different pipelines:

1. Native render. The model generates every pixel at the size you asked for. Seedance 2.5 does this up to 1080p, and its 1080p output is a 10-bit render. Wan 3.0 renders natively at every size it lists.
2. Latent refinement. MiniMax H3 Max generates at a native 768p. Its “1080p” option is a refinement pass inside the model’s latent space, done by the same endpoint and billed at exactly twice the 768p rate ($0.16 against $0.08 a second on fal’s regular pricing). It is not a 2K model, and nothing larger is on offer.
3. Render small, upscale outside. A lower tier like Seedance 2 Mini can be rendered at 480p and pushed through a separate super-resolution model. Volcengine’s MediaKit enhancer charges about CNY 2.5 per output minute up to 720p and CNY 5 up to 1080p at 24 fps. That is a fraction of a cent per second, against tens of cents for a native 1080p render.

For a viewer on a phone, an upscaled 480p clip can be perfectly acceptable. For anything with fine texture, small text or hair in motion, a native render is visibly better, and the price gap is the honest measure of that difference. When you compare models, ask which of the three pipelines produced the resolution on the label.

Audio now comes out of the same pass

The quiet revolution of this generation is that sound is no longer a separate product. Seedance 2.5 renders its own audio, including lip sync for speaking characters. MiniMax H3 Max renders a synchronised track in the same pass with no switch to turn it off. On my own test renders it came back as stereo AAC at 32 kHz. Wan 3.0 also generates audio, and fal charges the same per-second price with or without it.

That last detail matters. When audio is free at the margin, it stops being a feature and becomes part of the model’s output distribution. Footsteps land on the frame where the foot hits the floor because both come out of one joint generation, not two models stitched together after the fact. The flip side is control. You can describe the sound in the prompt, but you cannot yet mix it. Anyone who needs a clean dialogue stem still ends up replacing the track in an editor.

Thirty seconds in one take

Clip length is where the models diverge most:

• Seedance 2.5: any whole number of seconds from 4 to 30.
• Wan 3.0: 2 to 30 seconds through fal.
• MiniMax H3 Max and H3 Max Turbo: 5 to 15 seconds.
• Kling 2.5 Turbo Pro, still a favourite for complex motion: fixed 5 or 10 seconds at 1080p.

A 30-second single take is a different kind of object from a 5-second clip. The model has to keep a character’s face, clothing and the geometry of the room stable across 720 frames at 24 fps. In older models that drift was hidden by keeping clips short and cutting often. Seedance 2 already learned to cut between shots inside one clip. Seedance 2.5 and Wan 3.0 push the other way, holding a continuous camera for half a minute. The cost of a failed take scales with it too, which is another reason the draft-low, render-high habit has become standard.

Reference trays are the new prompt

Text prompts are giving way to multimodal “reference” input, where you upload assets and cite them in the prompt by position. The Seedance 2 family documents a tray of up to 9 images, 3 videos and 3 audio files, 12 in total. The prompt refers to them as @Image1, @Video1 and @Audio1. Seedance 2.5’s tray can technically hold 50 files. H3 Max uses its own syntax (“Image 1”) on a separate reference endpoint.

The billing reveals how the providers think about it:

• Ark bills a Seedance 2.5 request that cites reference video on input seconds plus output seconds, at a lower token rate (CNY 42 per million up to 720p, 46 at 1080p). Your reference clip is processed like video the model has to watch.
• fal bills H3 Max reference assets as tokens: the first 4,096 are free, then $0.02 per thousand, where a 1024 × 1024 image counts as 1K. Four ordinary stills cost nothing extra, and a tray of 4K photos costs real money.
• Wan 3.0’s reference endpoint takes clips and sound as well as stills, but caps the clips and sound at 15 seconds in total and bills a reference clip’s own length on top of the output.

In other words, the reference tray is becoming context in the same sense as an LLM’s context window. It is metered, capped and priced per token.

Speed is now a product tier

The most surprising number in my logs is not a price but a timer. fal post-trained its own variant of MiniMax H3 and runs it on its own inference stack. I measured H3 Max at about 7.9 seconds of inference for a 10-second take. H3 Max Turbo took about 1.5 seconds for a 5-second take, at half the per-second price at every size. For comparison, a 480p Seedance 2 Mini render took around 135 seconds, plus roughly as long again for an upscale pass.

When generation is faster than playback, the interaction model changes. You stop “submitting jobs” and start iterating the way you would with an image model. The trade-off is in the catalogue: Turbo is served for text-to-video and image-to-video only, with no reference endpoint. Anything that needs a tray of reference material still goes to the slower sibling. If you want to see how the two H3 tiers behave on the same prompt, a browser workbench for MiniMax H3 lets you switch between both tiers on the same prompt. It is an independent tool, not MiniMax’s own.

A quick cost comparison

Upstream list prices per second in early October 2026, before any reseller margin:

fal ran a 40% promotion on both H3 tiers until 15 October, so anyone who benchmarked costs in early October saw lower numbers than the regular rates above.

What to take from it

Three conclusions hold whichever model you favour:

1. Price by pixels, not by clip. Tokens track width × height × fps × seconds, so resolution is the biggest lever on your bill. Draft at 480p and decide what earns a native 1080p render.
2. Read the pipeline behind the label. Native, latent-refined and upscaled 1080p are three different products at three different costs.
3. Treat references like context. They are metered, capped and increasingly where the real control over a shot lives.

If you want to try Seedance 2.5 itself without wiring up an Ark account, Dolai runs it in the browser for text-to-video and image-to-video with sound. Whatever you use, keep a spreadsheet of what each test cost. In this generation of models, the arithmetic is half the craft.