Cost Per Usable Second Is the Only Video Price That Matters

A Chinese tech outlet made a one-minute GTA 6-style clip of an old woman storming a scrapyard and paid 5.4 yuan for it. That works out to about nine fen a second, or roughly one US cent per second of 720p video.

The clip is not the interesting part. The price is, because it changes what a reasonable workflow looks like. When a second of video costs less than a stick of gum, regenerating a shot stops being a budget decision and becomes an editorial one, and that is a different craft.

Vidu Q4 preview comes from Shengshu Technology and opened to creators in early October 2026. The QbitAI piece is launch-week coverage with real numbers in it, so it is worth reading for those. It is also promotional in tone, with an invite code and a first-tier claim at the end. Treat the price data as measured and the enthusiasm as marketing.

The number that should be on your spreadsheet

Price per second is a list price. It is not a cost.

The cost that matters is price per second divided by acceptance rate, which gives you cost per usable second. If one take in three survives your review, nine fen becomes twenty-seven fen. If one in six survives, it becomes fifty-four fen. A pricier model only wins when its acceptance rate improves by more than its price multiple, and that is a high bar: three times the price has to land more than three times as many usable takes to break even.

Nobody publishes acceptance rates, because acceptance depends on your subject. Faces hold differently than products, products hold differently than abstract effects, and a model that nails one will fumble another. So the only way to get the number is to run ten takes on your own prompt with your own references and count.

Ten takes is a deliberate choice of sample size. Three takes tell you almost nothing about a rate. Ten is enough to distinguish "about half" from "about a sixth" without spending an afternoon.

What the cheap tier actually offers

The Q4 preview is not a stripped-down model with a discount attached. The published capabilities are a 16-second maximum per generation, camera moves (tracking, orbit, push and pull, crane), first-person FPV framing, resolution switching between wide, medium and close shots inside a single generation, and 4K output through both the API and the product.

The reference budget is the more interesting spec: up to 15 reference images plus 3 reference audio clips, with each asset placed into its own slot in the prompt. Character, costume, prop and location can be separate references, and the prompt describes how they relate.

That changes the failure you spend your time on. Identity consistency stops being a prompting problem and becomes an asset problem. If you have eight clean angles of the same jacket and three seconds of the right voice, the model has something to hold onto. If you have one compressed phone photo, you have a lottery ticket with better odds than last year and the same fundamental problem.

The audio references are the underrated part. Three clips of a voice, used across scenes and emotions, is the difference between a character who sounds like a person and a character who sounds like a different person every cut. A voice is an identity asset exactly the way a face is, and almost nobody prepares it in advance.

Where the cheap price gets spent

The QbitAI test cases are worth listing because they show what the low price is buying: a two-person over-the-shoulder dialogue in English, a comic fight scene, a science-fiction reveal where a structure lights up and the light spills onto a spacesuit and a cloud sea, a coffee ad, and a group of user-made clips in ink-painting and cartoon styles.

The sci-fi shot is the one I would point at. Making an explosion is easy now. Making the light from the explosion land on the surrounding objects, so the suit and the railing and the clouds all shift colour, is the thing that makes an effect read as part of a scene instead of a sticker. That is not a feature you can prompt your way into on a model that does not do it.

The multi-character dialogue is the one I would hedge on. A single generated take holding two people in conversation is a genuinely hard case, and the article shows one good result, not a rate. One good result is a demo. Ten good results out of twelve is a workflow.

Two ways the math goes the wrong way

First trap: the nine fen figure is promotional. The API list price in the same article is about 0.6 yuan per second at 720p and 0.75 at 1080p, with tiering, and the cheap SaaS rate is described as a two-month launch promotion. That is roughly a factor of seven. A pipeline designed around 0.09 yuan per second becomes a different pipeline at 0.6. If you are building cost assumptions now, build them on the list price and treat the promotion as a bonus, or you will discover the real cost at renewal with footage half shot.

Second trap: resolution. The cheap number is 720p. Going to 1080p moves the per-second price by a quarter before any retries, and the launch coverage gives no 4K per-second rate, so treat 4K as a further step up rather than a small one. If your deliverable is a vertical social clip that will be watched on a phone, 720p with a good composition beats 4K with a bad one, and the cost difference pays for several more attempts at getting the composition right.

The trade the cheap tier is actually making

Cheap generation shifts the work from writing to selecting, and selecting has its own discipline.

When a take costs real money, the rational move is to spend ten minutes perfecting a prompt. When a take costs a cent, the rational move is to describe the shot in three lines and generate eight versions with deliberate variation: one camera angle apart, one beat of timing apart, one framing tighter. You are buying coverage rather than precision.

Coverage only works if the variations are planned. Eight random takes of the same sentence is not coverage, it is a slot machine with a receipt. The useful version is a shot list written before you generate: same subject, camera at eye level / low / three-quarter, action at second two versus second five, wide versus close. Then you pick from a set you designed instead of a set you hoped for.

There is also a structural saving that the price table hides. Sixteen seconds in one generation, with the ability to cut between wide, medium and close inside it, means fewer clips per finished minute than a five-second-per-clip model requires. Fewer clips means fewer joins, and joins are where continuity dies. A cheap model that produces four clips for a minute of screen time can beat a marginally better model that needs twelve, before you count a single retry.

What I would still not hand to the cheap tier

A shot whose text has to be exact. A character whose face is the product. Anything where a client will pause on a frame.

The promotional cases in the source article are broad gestures: a fight, a light reveal, a coffee pour. Those tolerate small errors because motion carries them. Typography and faces do not, and both are where a low-cost model shows its price.

Run the test before you commit

The procedure is small enough to do in an afternoon.

Pick three shots from your actual next job: one face, one product, one effect. Write a normal prompt for each, attach the references you really have, and generate ten takes per shot at the resolution you plan to deliver. Count how many you would ship without a fix. Ten takes times the clip length times the list price, divided by the number you would ship, is your real cost per usable clip. Divide that by the clip length and you have cost per usable second, which is the version that compares across models and shot lengths.

Compare that number across the two models you are actually considering, not across a leaderboard. If the cheap model wins on cost per usable second and you can live with its ceiling, use it and spend the savings on coverage. If it loses, you now have a number instead of an opinion, and the number survives the next launch cycle better than the enthusiasm does.