What a finished minute of AI video really costs

The real AI video generation cost per minute is the listed price per second of output, multiplied by the number of attempts it takes to get each usable shot, plus the extras you pay for along the way (upscales, extensions, reframes, variations for the client) and the finishing work that turns a pile of clips into a minute someone can watch. A rate card tells you what one generation costs. A finished minute usually needs many generations, most of which you throw away, and then several hours of editing, colour and sound. To budget honestly, measure your own retry ratio on a short test, price every extra, and treat finishing as its own line rather than an afterthought.

Prices for generation services change often and differ by model, resolution and plan. Nothing below quotes a vendor's number; check each vendor's current pricing and plug it into the method.

Why the per-second rate misleads

Generation services usually express cost in credits or currency per second, per clip or per generation. That number is accurate, but it describes a single attempt at a single shot. A finished minute is different in three ways.

• It is made of many shots. A minute of edited video often contains a dozen or more cuts, and each shot is generated separately, usually as a short clip.
• Most attempts are not kept. A hand goes wrong, a face drifts, the camera does something you did not ask for, or the lighting does not match the previous shot. Each fix is another paid attempt.
• Generated seconds are not used seconds. You generate a few seconds and use the best part of it. Trimming the start and end of every clip means you pay for more seconds than end up on screen.

So the per-second rate is a floor, not an estimate. The useful question is how many paid seconds it takes to produce one second you keep.

The four layers of a finished minute

Break the cost into four layers and estimate each separately.

1. Kept seconds

Sixty seconds of final footage at the listed price, the only layer the rate card describes. It is usually the smallest number on the sheet.

2. Retries and failed takes

This is usually the largest layer. Count it as an attempts-per-keeper ratio: how many generations, on average, before a shot is usable. Simple shots (a landscape, an abstract texture, a slow push-in on an object) may need few attempts. Shots with people, hands, text, fast motion or continuity with the previous shot tend to need many more.

Also check the vendor's policy on failed jobs. Depending on the service, a generation that errors out or is blocked by a content filter may or may not consume credits. Read the terms rather than assuming either way.

3. Overgeneration and extras

• Handles. You generate longer than you use so you have room to cut. Those extra seconds are paid.
• Variations. If the client sees two or three options per shot, multiply accordingly.
• Resolution. Many teams iterate at a lower resolution and pay again for the final at a higher one. Count both.
• Extensions, upscales and reframes. These are often separate charges, and a vertical version for social can mean another round.
• Revision rounds. Client notes on a generated shot frequently mean generating it again, not adjusting it.

4. Finishing

Generated clips rarely cut together cleanly as delivered. Shots from different generations differ in colour and contrast, the motion may need trimming, there is no real sound design, and titles and captions still need making. Finishing covers editing, colour matching, sound, titles, captions, the export and a final check. It is mostly paid in time.

Finishing is also where you can cut layer 2. Many problems that tempt you to generate again are editing problems: a clip that runs too long, a shot that is slightly too warm, a framing that crops the subject. Our guide to editing AI video without re-rolling goes through which fixes belong on a timeline and which really need a new generation.

A worked example in units, not prices

To show the arithmetic without inventing vendor prices, count everything in "generated seconds" and convert at the end using your vendor's current rate.

Say a finished minute needs fifteen shots of about four seconds each.

1. Kept seconds: 60.
2. Handles: you generate six seconds to use four, so each attempt is six paid seconds.
3. Retries: your test showed an average of five attempts per kept shot. That is fifteen shots times five attempts times six seconds, which is 450 generated seconds.
4. Variations: the client sees two options for five hero shots. Add five shots times five attempts times six seconds, another 150 seconds.
5. Revisions: you expect one round of notes affecting a third of the shots. Add five shots times five attempts times six seconds, another 150 seconds.
6. Extras: upscale the fifteen final shots once each, priced per the vendor's upscale rate.

That is 750 generated seconds, plus upscales, for 60 seconds on screen: a ratio of about 12.5 paid seconds per kept second, before finishing. Your own ratio could be much lower or much higher. The point is that the multiplier, not the rate, decides the budget, and you only learn the multiplier by testing.

Then add finishing as hours. Time yourself finishing the test shots and scale up.

How to measure your own ratio

A short pilot gives you better numbers than any guide.

1. Pick three or four representative shots, including at least one with people or motion that is hard to generate.
2. Generate until each is usable, and log every attempt, its length and its resolution.
3. Note why each attempt failed. Anatomy, continuity, camera behaviour, lighting or prompt misunderstanding. This tells you which shot types will be expensive.
4. Finish the pilot. Cut the kept shots together, match the colour, add rough sound and export. Time it.
5. Write down two numbers: generated seconds per kept second, and finishing hours per finished minute.

Repeat the pilot when you switch models or vendors.

Budgeting a project

With your two numbers in hand, the estimate is straightforward:

• Generation cost = kept seconds times your paid-to-kept ratio times the vendor's current rate, plus extras priced separately.
• Finishing cost = finished minutes times your finishing hours times your hourly rate, plus the software you finish in.
• Margin = a buffer for the shots that go badly. First projects with a new model deserve a larger one.

Then look at what the total does not show. Generation cost is variable: it rises in exactly the projects where the brief is vague or the client wants many options. Agree in advance how many variations and revision rounds are included, and say in the quote that further rounds are billed, because each round is a new set of generations. Our explainer on AI video credits and paying per generation lists the questions to ask a credit-based vendor before you commit a production to it.

Where finishing software fits

Finishing is where a generated minute becomes a delivered one, and it is also the layer where software cost is easiest to predict. A flat-rate editor costs the same whether you fix one shot or fifty, which makes it sensible to push as many fixes as possible from regeneration into editing.

cinemaOS is one option for that layer on a Mac. It is not a text-to-video generator, so the generation lines on your sheet stay with whichever service you use. What it covers is the finishing: cutting the clips in the Sequencer, grading and matching shots in the Finisher, sound in Sound mode, titles and design in Atelier, and subtitles. A quick primary pass on mismatched generated clips is shown in the one-minute grade tutorial. Its J.A.R.E.D. assistant runs on the Mac with no generative credits, so using it adds nothing to the meter. Projects are local .mntm files with recovery, so your finished edit stays editable rather than ending as a single flattened file. The plan is $9.99 USD per month excluding VAT, with a 60-day free trial that needs no card and has no automatic charge.

The limits matter for AI work in particular. cinemaOS runs on Apple silicon Macs only today, with Windows planned for November 2027. It does not do lip sync, so a generated character whose mouth does not match new dialogue still needs another tool or a new generation. And it is not a full After Effects replacement, so heavy compositing on generated plates may need a companion app. Other editors, such as DaVinci Resolve, Premiere Pro or Final Cut Pro, finish generated footage too; the useful comparison is what each costs you over a year of projects, which our five-year cost worksheet helps you work out.

FAQ

Is AI video cheaper than shooting?

Sometimes. For footage that would be expensive or impossible to shoot, such as an aerial over a place you cannot reach or a concept visual for a pitch, generation can cost far less than a shoot. For shots with people, dialogue and continuity, the retry ratio and finishing time can close the gap quickly. Run a pilot and compare the full cost, not the rate card against a day rate.

Why does my AI video cost so much more than the listed price?

Because the listed price is per attempt, and a finished minute needs many attempts plus handles, variations, revisions and extras. Measure your paid seconds per kept second on a short test; that multiplier explains most of the difference.

How can I lower the cost per finished minute?

Iterate at lower resolution and pay for the high-resolution version only once a shot works. Agree the number of variations and revision rounds up front. And fix timing, colour, framing and sound in an editor instead of regenerating, keeping new generations for shots whose content is actually wrong.