L'article ci-dessous est publié en anglais. Chaque chiffre qu'il contient est le chiffre publié par l'éditeur lui-même, avec la date à laquelle nous l'avons lu.
A feature grid assumes the eight tools profiled here are the same kind of purchase. Three of them are not. Start instead from what you are making, and most of the field stops applying before you have read a single price.
Three jobs, and only one is what people mean by generating video
The eight profiles in AI Video and AI Avatars cover three different products. Two make a presenter say something you wrote. Five make a shot from a prompt, and one of those five no longer exists. One does not generate anything at all: it edits footage you already have, by editing the transcript. Deciding which of the three you need removes most of the field immediately.
- A presenter explaining something. Synthesia and HeyGen put an avatar on screen reading your script.
- A shot from a prompt. Runway, Pika, Luma Dream Machine and Google Veo generate motion that did not exist. OpenAI Sora belonged here and was withdrawn, which the last section returns to.
- Footage you already have. Descript is an editor rather than a generator, and its own profile says so: you edit the transcript and the video follows.
Ranking those three against each other on output quality would mean nothing, which is why our AI Avatars hub carries the same warning about its two members. The comparison that does work across all of them is cost, because a budget is a question you can put to any of them.
Three ways of charging, and the answer is not a number
Ask what an hour of use costs and the eight profiles do not answer in the same currency. Four print a monthly subscription. One prints a rate per second of finished output. Two print no entry price at all. That is three different shapes of answer before anyone argues about which is cheaper.
| Tool | Entry price, as the vendor prints it | Shape |
|---|---|---|
| Runway | $15/mo, Standard | per month |
| Descript | $24/mo per person, Hobbyist | per month, per person |
| Synthesia | $29/mo | per month |
| Luma Dream Machine | $30/mo, Plus | per month |
| Google Veo | $0.05 per second (720p) | per second of output |
| HeyGen | not published on this profile | none |
| Pika | not published on this profile | none |
Every figure is the vendor's own, and each profile names the page it came from and the day a human opened it.
A rate per second has no floor and a subscription has one
This is the structural difference, and it decides more than the size of the numbers does. A subscription is a standing cost: you pay it in a month where you make nothing. A rate per second of output is only charged against output, so a month with no work costs nothing and a heavy month has no ceiling. Neither is better. They fit opposite patterns of use.
If the video work is occasional, a per second rate is the shape that matches it, and among the seven tools still available Google Veo is the only one that prints one. Sora printed the other, and both are API rates rather than consumer plans, which is the pattern rather than a coincidence: a rate per second is what you quote to a developer building on top of you. If the work is continuous, the subscription's ceiling is the whole attraction, and four tools offer that.
The catch is on the same page as the answer. Google's own pricing page prints Not available in the Veo free tier row, so on the surface we read there is no way to find out whether it fits without being billed for finding out. That scope is the whole of the claim: Veo's profile says we did not check Google's consumer products for a separate way in, and neither should this sentence. Our AI video ranking places Veo last for exactly that reason and says in its own note that the ranking would look very different on output quality.
Two of the eight publish no entry price, and the reasons are not the same
An empty entry price is not a statement about what a vendor charges. It records that we did not establish a figure, and the two cases here were stopped by different things. Reading them as though they meant the same thing is the mistake that page exists to prevent.
HeyGen prints a figure against its cheapest paid plan and offers both a monthly and a yearly cadence, with nothing in the document tying the figure to one of them. We read that page twice, the second time with an explicit English preference, and both times it served its Vietnamese edition. Pika's page loads with the yearly option already selected, so the month to month figure is not legible on it.
Both of those are facts about a page and about our reading of it. Neither is a fact about the product. Three kinds of absence works through why we leave the row off rather than fill it.
Where our own ranking stops answering your question
Our AI video ranking sorts on how much you can produce before you have to pay, and it states that axis at the top because it decides the order. That is the right question while you are still deciding whether to pay at all. It is not the question in this guide, and the two do not produce the same order.
Google Veo ranks last there and is the one tool here whose bill scales with what you make. Synthesia ranks first there because its free tier is measured in finished video, which is the unit you actually deliver in. Same eight profiles, same figures, two different questions, and no contradiction between them as long as each page says which one it is answering.
One of the eight is not an option at all. OpenAI ended the Sora apps on April 26, 2026 and has published September 24, 2026 as the date the API follows. Its profile stays up because the question people arrive with is whether it is still available, and a deleted page answers that with a 404.
What we will not tell you
We will not rank these seven live tools on price, because they do not share a unit. A rate per second and a monthly subscription are not two points on one scale, and two of the seven publish nothing to place at all. A table that sorted them anyway would be inventing an order, and the ordering would look authoritative in exactly the way a wrong one does.
What the profiles above let you do instead is settle the job first, then the shape of the bill, and only then compare figures that share a unit. The allowances that decide the trial rather than the bill are covered in why credits, minutes and clips do not convert.