Tries per Usable Shot: The Number That Decides What AI Video Costs

The ratio

Tries per usable shot is simple arithmetic: the number of generations you ran, divided by the number of shots you would put in front of a client. Two models can charge the same per generation and land very far apart on this number. That gap is the real difference in price, and it never appears on a launch page.

It also hides inside a habit every AI filmmaker knows. You slow the camera down and the product label changes. You fix the label and the background moves. Each fix costs a generation, and each generation costs credits and the time it takes to watch it.

Why the price per generation misleads

Take an illustrative case, not a price list: say one generation of a 10-second shot costs $3 in credits. If a model gives you a usable take every second try, a usable 10 seconds costs $6. If it needs eight tries, the same 10 seconds costs $24, before anyone counts the time spent reviewing the seven rejects.

The credit rules make the point sharper. Kling’s credit policy refunds generations that fail technically, but a take you reject because it looks wrong is still charged, as Kingy AI notes. The model is paid for every try. Only you know how many of them were usable. The same logic applied to choosing a model, shot by shot.

Three tests that move the number

Before a new model touches client work, we put it through three tests, chosen because they are where AI video breaks most often on brand work.

Product lock. Does a labeled product survive a 20 to 30 second orbit or push-in, letter for letter? Brands notice a misspelled label before they notice anything else.

Keyframe obedience. Set five to ten beats on the timeline. Does each one land within about a second of where you asked, with nothing invented in between? Early testers of Kling 4.0 Flash saw the model add a gesture and a jacket nobody asked for, the kind of invention this test catches.

Lip sync in two languages. Run the same two lines in English and in Québécois French. Does the mouth follow syllable for syllable, with the same face in both versions? For a bilingual market, a model that only syncs English does half the job.

Reading the score

Count the tries for each test. One or two tries per usable shot: the model is ready for client work. Three to five: usable, with extra budget set aside. Six or more: keep it for the specific shots it does well and use something else for the rest.

A model that passes all three tests saves money. A model that passes one is a tool for specific shots, and that is still useful, as long as you know which shots.

Fix one thing, check three

The rule we hold to: after any fix, check all three tests again. A slower camera can break the label; a corrected label can break the lip sync. Measuring only the thing you just fixed is how a ratio of two quietly becomes a ratio of six.

What it means if you buy video

Per-second pricing only holds when the studio knows its ratio. Ask to see a test on a product like yours, and ask how many tries a usable shot took. A studio that tracks the number can price with confidence. A studio that does not is pricing a guess. How Kling 4.0 looks on day one.