What you are making decides how you pay

Bài viết bên dưới được đăng bằng tiếng Anh. Mọi con số trong đó là con số do chính nhà cung cấp công bố, kèm ngày chúng tôi đọc được.

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 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 OpenAI Sora belonged here and was withdrawn, which the last section returns to.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. Six print a monthly subscription. One prints a rate per second of finished output. That is two different shapes of answer before anyone argues about which is cheaper.

Every figure is the vendor's own, and each profile names the page it came from and the day a human opened it. Where the entry price and the terms pull in different directions, the review reads them together: the Pika review puts the $10 Starter against the commercial license that only arrives at $35 and the training clause in the same paragraph as the ownership sentence.

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.

When a row is left off the table

An empty entry price is not a statement about what a vendor charges. It records that we did not establish a figure, and the reason it is empty is specific to the page. Reading an absence as a fact about the product is the mistake this section exists to prevent. Today every profile in this guide carries a figure, and that was not true in July: Pika's page then loaded with the yearly option already selected, so the month to month figure was not legible, and the profile said so rather than guess. The page changed, and the row filled with the monthly figure it now prints.

That is a fact about a page and about our reading of it, not a fact about the product. Three kinds of absence works through why we leave a 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 shut down the Sora API on September 24, 2026. 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. 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.