AI video looks expensive because one cost dominates: the video model. Everything else in a film is a rounding error. Understanding the split explains both how pricing works and where to save. This article gives orders of magnitude rather than exact figures, because provider prices change monthly; the ratios between them are what stays stable.

Provider costs (orders of magnitude, 2026)

  • Video, fast tier: roughly 0.05–0.12 dollars per second of output. These models produce a five- or ten-second clip in one to three minutes and are good enough for drafts, social clips and drawn styles.
  • Video, cinematic tier: roughly 0.15–0.40 dollars per second. Slower, richer motion, better physics, and a tendency to add photorealistic texture that drawn styles do not want.
  • Image (character sheet, keyframe): roughly 0.03–0.06 dollars each. A film needs one sheet per character and one keyframe per scene.
  • Speech: roughly 0.10 dollars per thousand characters. A four-minute film has under two thousand characters of narration and dialogue combined.
  • Lip sync: priced per second of the clip being synced, typically in the fast-video range, and only for scenes with a spoken line.
  • Music: a few cents per minute.
  • Screenplay generation: cents per film on a flash-class language model, tens of cents on a frontier model. The frontier model is worth it for the director stage and nowhere else.

A worked example

Take a four-minute film with thirty scenes of eight seconds, four characters, twelve spoken lines and narration in every scene. On the fast tier: video is 240 seconds at about 0.08 dollars, so around 19 dollars. Sheets and keyframes are 34 images at 0.05, under 2 dollars. Speech, music and direction together are under 1 dollar. Lip sync on twelve eight-second clips adds 5–8 dollars. The film costs the provider side 27–30 dollars, of which video and lip sync are over 90 percent. On the cinematic tier the video line alone becomes 50–90 dollars and everything else is unchanged.

From dollars to tokens

A token wallet expresses every operation in one unit so that the person never sees seven currencies. The honest way to set prices is a single formula: token price = real cost × margin multiplier, at the rate a person actually pays for tokens, anchored to the most discounted plan so that the margin never goes negative for anyone. If the top plan pays 0.8 of list price, the formula uses 0.8, and every other plan simply earns more.

The formula must be applied per operation, not per film, because operations are what get charged and refunded. A film is thirty-odd charges; if one fails, that one is refunded. Pricing per film would hide failures inside an average and make refunds impossible to reason about.

What a fair margin looks like

Around 40 percent on high-volume creative work covers infrastructure, refunds, support and the occasional provider outage. Margins of several hundred percent were common when "generation" meant one clip; they cannot survive a pipeline that makes thirty generations per film, because the person sees the total. A competitor that prices a four-minute film at 30–40 dollars is a competitor that has done this arithmetic; a platform charging 300 for the same film is one that has not.

Where the savings are

  1. Generate short first. Iterate on a one-minute cut of the opening, fix the cast and the style, then commit to the full length. Most creative decisions are visible in the first minute.
  2. Use the fast tier for animatics and drafts, and reserve the cinematic tier for the final pass on scenes that need it. Mixing tiers within one film is fine; the assembly step normalises resolution.
  3. Compose keyframes before animating. A failed frame costs cents; a failed clip costs dollars. Reviewing thirty keyframes takes two minutes and catches most composition errors.
  4. Skip lip sync on wide shots. Mouths are not visible; the line can play in the audio track with no loss.
  5. Reuse cast sheets across films. The characters are drawn once.

What subscriptions have to cover

A plan sells a monthly token allowance for a fixed price. The solvency rule is simple: if the whole allowance is spent on the most expensive operation, it must still cost the provider less than the plan price. Do this arithmetic before publishing prices, and redo it whenever a provider changes a rate. A plan that loses money on its heaviest users will attract exactly those users.

Frequently asked

Why not charge in dollars directly? Dollars expose provider prices and turn every rate change into an argument. Tokens let the platform absorb small changes and only reprice when the underlying cost shifts materially.

Why are failed jobs refunded automatically? Because the provider does not charge for them, so neither should the platform. Refunds must be idempotent so that a retried refund never pays twice.

Do longer films get cheaper per second? No. Video pricing is linear in seconds; there is no bulk discount from providers, and a platform should not invent one it cannot fund.