Every AI video tool prices itself in credits, and no tool prices itself in minutes. That is awkward, because a minute is the unit you actually need. A thirty-second product clip, a two-minute explainer, a short course module — those are the things being budgeted for, and none of them appear on a pricing page.
So we did the conversion. Two of the three services below publish enough detail to do it honestly. The third does not publish it at all.
One caveat before the numbers: these are not the same product. HeyGen makes a presenter say a script. Luma generates a scene from a prompt. Comparing their per-minute cost is useful for budgeting, not for deciding which is better at a job only one of them does.
HeyGen: the rates are published, and the range is wide
HeyGen deserves credit here. Its help centre lists the exact credit cost of every feature, per minute of output, which is more than most of this category manages.
The Creator plan is $29 a month and includes 600 credits. That works out at just under 4.9 cents a credit, which turns the published rates into this:
| What you are making | Credits per minute | Cost per minute | Minutes on a month of Creator |
|---|---|---|---|
| Avatar III, video look | 4 | $0.19 | 150 |
| Avatar IV, photo look | 16 | $0.77 | 37 |
| Avatar IV, video look | 31 | $1.50 | 19 |
| Avatar V | 48 | $2.32 | 12 |
| Video Agent, standard | 80 | $3.87 | 7 |
| Video Agent, Seedance mode | 240 | $11.59 | 2 |
The spread is the point. The same $29 buys either two and a half hours of avatar video on the older engine, or two and a half minutes on the newest one. Nothing on the pricing page tells you that, because the pricing page sells you the plan, not the rate.
Some extras stack on top rather than replacing the base cost. Custom Expressive Motion is 40 credits a minute “in addition to the Avatar IV generation cost”. Upscaling to 4K is 2 credits per second — 120 a minute, four times what the Avatar IV video itself cost to make.
Two things to know about topping up. Creator users cannot buy more credits at all; the only route is upgrading to Pro. Pro users cannot buy packs either, and instead move up a tier — the tiers run from 1,000 credits at $49 a month to 100,000 at $4,300. Only Business customers, at $149 a month plus $20 per seat, can buy one-off credits, at 5 cents each.
The free plan is three videos a month, capped at one minute each, with a watermark and no 1080p export. It is a demonstration, not a workable free tier.
Luma: the most honest rate card in the category, and the most sobering
Luma publishes a rate card that runs to dozens of rows — every model, every resolution, with and without audio, priced per second. It is genuinely the clearest document of its kind we have read. It is also the one most likely to change your plans.
There is no free tier. Individual plans start at Plus, $30 a month for 10,000 credits — three tenths of a cent each. Here is what a finished minute costs at that rate, and how much of the monthly allowance it consumes:
| Model and setting | Credits per second | Cost per minute | Minutes on a month of Plus |
|---|---|---|---|
| Ray3.2, 720p | 20 | $3.60 | 8 min 20 s |
| Kling 3.0, 1080p with audio | 49 | $8.82 | 3 min 24 s |
| Ray3.2, 1080p | 80 | $14.40 | 2 min 5 s |
| Veo 3.1, with audio | 140 | $25.20 | 1 min 11 s |
| Seedance 2.5, 1080p with audio | 191 | $34.38 | 52 s |
A $30 monthly plan buys about fifty seconds of the newest model at full quality. That is not a criticism of Luma, which has published the arithmetic for anyone willing to do it. It is a description of what frontier video generation currently costs.
Volume helps, but less than you might hope. Ultra is $300 a month for 150,000 credits, which brings the credit price down from 0.3 cents to 0.2. The same Ray3.2 minute at 1080p falls from $14.40 to $9.60 — a third off, not an order of magnitude.
Read the footnotes too. On Ray3.2, HDR doubles the cost and HDR with EXR triples it, and generating without keyframes adds 30 credits. Upscaling to 4K through Luma is 25 credits a second on top — another $4.50 a minute on Plus.
Leonardo: you have to sign up to find out
Leonardo sells four individual plans — Free, Essential at $12, Premium at $30 and Ultimate at $60, all excluding tax — measured in “fast tokens” per month. What it does not publish, anywhere we could find, is how many tokens a video generation consumes.
Its API documentation is explicit about where that information lives: the pricing calculator “is available in the API section of the Leonardo App”. You can see what a generation costs after you have an account, not before.
Two details are worth knowing before you compare plans. Unlimited video generation “at a relaxed pace” appears only at Ultimate, $60 a month, and only for a named list of models; the Premium plan’s unlimited allowance is images only. And on the free plan, the “Creations access” row reads Public. Private generations start at $12 a month. If you are trying the tool on anything you would rather not publish, that is the line to read first.
The cost nobody prices: the attempts that did not work
Every figure above is the cost of one generation. Prompted video rarely lands first time. If it takes you four attempts to get a usable five-second shot, you have paid four times, and the pricing page still says one.
This is the single biggest gap between a credit allowance and a production budget, and it is why the cheap-looking plans tend to disappoint. Budget for the number of attempts, not the number of finished seconds. On the figures above, a $30 Luma Plus month at 1080p is roughly two minutes of successful output — or thirty seconds of it, if three attempts in four get discarded.
The open-source route, and what it really costs
Two serious video models are downloadable. They are free in the sense that nobody invoices you, which is not the same as free.
HunyuanVideo publishes its own requirements: 60GB of GPU memory for 720p output, 45GB at the smaller setting, tested on a single 80GB card, with 80GB recommended. Its own benchmark table puts a 1280×720 clip of 129 frames at 1,904 seconds on one GPU. That is about thirty-two minutes of compute for roughly five seconds of video. Rent that class of card by the hour and the arithmetic stops looking free very quickly.
There is also a licence to read. The Tencent Hunyuan Community License opens in capitals: “THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA AND IS EXPRESSLY LIMITED TO THE TERRITORY, AS DEFINED BELOW.” The definition confirms it — “Territory” means “the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea”. If you are in any of those three places, this is not a model you have been licensed to use.
Wan is the more practical of the two on both counts. Versions 2.1 and 2.2 are released under the plain Apache License 2.0 — no territorial carve-out, no user-count trigger. And the 5B model, in Alibaba’s own description, “supports both text-to-video and image-to-video generation at 720P resolution with 24fps and can also run on consumer-grade graphics cards like 4090”, with the documentation putting the requirement at 24GB of VRAM.
A 4090 is a real purchase, but it is a one-off against an unlimited number of attempts. For anyone generating video steadily, at 720p, who is comfortable running things locally, that is the only option on this page where the fifth retry costs nothing.
What we checked
HeyGen’s pricing page and credit documentation, Luma’s pricing page and rate card, Leonardo’s pricing page and API pricing FAQ, and the HunyuanVideo and Wan repositories and licence files were all read on 4 November 2026 and quoted directly. All per-minute and per-month figures are our own arithmetic from the vendors’ published credit rates and plan prices, shown so you can check them. Prices are in US dollars and exclude tax. We have not tested any of these tools, and nothing here describes output quality.
For the rights question — whether you may actually use what these tools make — see The best AI tools for images, video and voice, and whether you may actually use what they make. For how credit systems behave when they empty, see When the credits run out, what actually stops?