A price per generation only means something once you know what a generation is. On this platform one video generation is anywhere from one second to sixty, at a short side between 480 and 2,160 pixels, with or without a native audio track — and each of those three choices moves the bill more than the choice of model does. So rather than publish an average, here is the whole table: every video lane we run, what it accepts, which wallet pays for it, and what a real clip actually costs.
Three variables set the cost of a clip and none of them is the model's name. Length is the first: every hosted lane here is linear in seconds, so fifteen seconds costs two and a half times what six costs on the same endpoint. Resolution is the second, and it is the violent one — Seedance 2.0 Mini's own resolution steps are ×2 at 720p, ×5 at 1080p and ×10 at 4K over its 480p rate, so the same fifteen seconds is ten times dearer at the top of its picker than at the bottom. Sound is the third, and it is free on one model and doubles the bill on another.
That is why a headline number for "one AI video" tells you almost nothing. It is also why this page is mostly tables. Every figure below comes out of the same manifests the composer quotes from and the same code that bills the job when it finishes, so the number on this page and the number on your receipt are computed by one formula.
Note: Two wallets, and they do not convert into each other. ⚡ green refills on a clock and pays only for models running on our own GPUs. ✦ gold is bought with money or granted by a plan, and pays for the hosted endpoints — Kling, Veo, Seedance, Grok, FLUX — which invoice us per call. Signing up grants zero gold.
LTX 2.3 is the workhorse and it is billed per started five-second block rather than per second, because that is how the graph renders: a six-second clip occupies two blocks of GPU time and costs two. The block rate starts at ⚡6 and is quoted at a 720p short side. Resolution scales GPU time roughly with pixel count, so the manifest carries a multiplier — ×0.75 at 480p, ×1.5 at 1080p — and the client quote and the completion charge read that same multiplier.
| Clip length | Five-second blocks | 480p | 720p | 1080p |
|---|---|---|---|---|
| 4 seconds | 1 | from ⚡4.5 | from ⚡6 | from ⚡9 |
| 8 seconds | 2 | from ⚡9 | from ⚡12 | from ⚡18 |
| 16 seconds | 4 | from ⚡18 | from ⚡24 | from ⚡36 |
The free plan clamps every video render to a 480 px short side, and the clamp is applied on the server before the price is worked out — deliberately, so nobody is billed for pixels they did not receive. That makes the leftmost column the one a free account sees. ⚡20 a day against ⚡18 for sixteen seconds is the honest shape of free video here: one long clip a day, or roughly four short ones. Any paid plan removes the clamp and raises the daily cap.
MiniMax H3 Studio is the second own-GPU lane and it is billed per second instead — from ⚡2 a second, any whole length from 1 to 15, with a ×1.5 step at 1080p. It is also the only lane on our hardware that emits its own audio, and that audio costs nothing extra because the picture and the sound come out of the same latent. Its length ceiling is not a policy but a memory budget: 15 seconds at 720p has rendered on our card, 15 seconds at 1080p exhausted it twice, so the composer greys the impossible combinations out rather than quoting a job that would fail.
What it costs: Every own-GPU figure on this page is a floor rather than a quote. Prices on our own hardware are set at runtime, can carry a demand multiplier when the render queue is deep, and if the local lane is unavailable the same model is served over a paid API and billed in gold instead. The composer shows the real number on the Generate button before you press it.
Three own-GPU lanes take footage you already have rather than a prompt, and they re-render every frame of it, so they are billed per second of source with no block to round to. Video Upscale bills from ⚡10 a source second at 720p, ×1.5 at 1080p and ×4 at 4K — and its length ceiling falls as the resolution rises, to 60 seconds at 720p, 40 at 1080p and 10 at 4K. Watermark Remove and Subtitle Remove bill from ⚡30 a source second, and their output short side is capped at 544 px with nothing upscaled: they clean a clip, they do not enlarge one. All three open at Ultra.
Everything below runs on somebody else's hardware and invoices us per call, which is why it is paid in gold and opens at the Starter plan. The lengths and resolutions are each endpoint's real envelope rather than a menu, and the price column is one clip computed with the pricing formula the composer uses.
| Model | Clip length | Resolutions | One clip |
|---|---|---|---|
| MiniMax H3 | 3–15 s | 480p, 720p | ✦9 at 6 s, 480p |
| Grok 2 Imagine | 6 or 10 s | 480p, 720p | ✦11 at 6 s |
| Seedance 2.0 Mini | 4–15 s | 480p, 720p, 1080p, 4K | ✦13 at 6 s, 480p |
| Grok Imagine 1.5 | 1–15 s | 480p, 720p, 1080p | ✦17 at 6 s, 480p |
| Kling 3.0 Motion Control | 3–30 s, follows the reference clip | 720p | ✦22 at 5 s |
| Kling 3.0 Pro | 3–15 s | 1080p | ✦23 at 6 s |
| Kling O1 Video Edit | 3–10 s, the source clip's length | 720p | ✦29 at 5 s |
| FLUX 3 Video | 5–20 s | 720p, 1080p | ✦29 at 5 s, 720p |
| Veo 3.1 | 4, 6 or 8 s | 720p, 1080p | ✦101 at 6 s |
Read the last column twice. At six seconds the spread is ✦9 to ✦101 — eleven times, for the same nominal product. Veo 3.1 is the dearest per second of anything on the site by a wide margin, and the gap is not our markup: every price here is derived from what the call costs us against the same floor, so a model that bills us more bills you more in the same proportion.
Gold arrives two ways: a plan grants it monthly, or you buy a bundle. The free plan grants none at all, which is the single most important thing on this page — no amount of the refilling green wallet reaches a hosted video model, at any length.
| Plan | Gold a month | Six-second Kling 3.0 Pro clips | Six-second Veo 3.1 clips |
|---|---|---|---|
| Free | none | 0 | 0 |
| Starter | ✦50 | 2 | 0 |
| Creator | ✦380 | 16 | 3 |
| Pro | ✦900 | 39 | 8 |
| Ultra | ✦10,000 | 434 | 99 |
Watch out: Green credits are never for sale. Buying credits buys gold, which pays for hosted models and queue priority — it does not raise the daily green allowance and it does not lift the free plan's 480 px clamp. Only the plan does that.
We are not going to quote anyone else's prices, because we do not bill their models and would be reciting a marketing page. What we can hand you is the set of questions that make two numbers comparable at all. Ask them of us too.
Open the video composer — The Generate button carries the quoted price before you press it. Nothing runs until you do.
The own-GPU LTX 2.3 lane, paid in green credits that refill on a clock. At a free account's clamped 480 px short side a four-second clip is about ⚡4.5 and a sixteen-second clip about ⚡18, against a ⚡20 daily cap. Own-GPU prices are set at runtime, so treat those as floors.
Because it bills us more per second. Every hosted price here is derived from what the provider charges us against one margin floor, so the ratio between two models on this site is close to the ratio between their two invoices. Veo's native audio track doubles its own price on top of that.
Yes, on the models we run ourselves — LTX 2.3 in its four modes and MiniMax H3 Studio — paid from the free plan's ⚡20-a-day green wallet and clamped to a 480 px short side. Hosted models are not reachable with green credits at any length.
On the hosted lanes, yes — they are linear in seconds. On our own LTX lane it is per started five-second block, and the picker offers 4, 8 and 16 seconds only, which land on one, two and four blocks. So the cost per second of output falls as the clip gets longer: 16 seconds costs four times what 4 seconds costs and gives you four times the footage, while a hosted lane charges per second either way.
It does not. The multiplier for a boosted queue lane is snapshotted when you submit, and the own-GPU demand multiplier is applied to the quote you see on the button. What can differ is the price tomorrow: own-GPU rates are admin-set at runtime.