The short answer: it depends on your keeper ratio
The cost of an AI video is the total generation spend for the project divided by the finished duration you kept, plus the hours you put in around it. That is the whole answer, and it explains why two creators making apparently identical videos report costs that differ by several times. Neither is lying. One of them is discarding most of what they generate and the other is not, and the discarded clips cost exactly the same as the kept ones.
So the number you actually need is not on a pricing page. It is a property of your own workflow, and it changes as you get better at prompting. A useful way to think about it is that you are not buying video, you are buying attempts. Skill in this medium shows up as a smaller number of attempts per usable second, which is also the only lever that reliably reduces cost without reducing quality.
Everything else in this piece is about measuring that properly, then converting it into a price you can defend to a client without guessing.
Cost Per Finished Second, the only figure worth tracking
Cost Per Finished Second, or CPFS, is the club's working metric for this. Take everything the project cost you and divide it by the number of seconds that survived into the final cut. Not the seconds you generated. The seconds a viewer will actually see.
The arithmetic is deliberately simple. If a generation costs C credits and you keep one in every K attempts, then each usable clip cost C multiplied by K. Divide that by the length of the usable clip in seconds and you have your CPFS in credits. Suppose a generation costs 10 credits and you keep one attempt in four: each keeper cost 40 credits, and if the keeper runs 5 seconds, that is 8 credits per finished second. Those figures are illustrative arithmetic rather than a benchmark, and your own C and K will be different.
- C is what one generation costs you at the model, resolution and duration you actually use. Read it live, since it varies by model and by settings.
- K is your keeper ratio, and it is the number almost nobody measures. It is also the number that moves most as your prompting improves.
- Finished seconds means what is in the export. A clip you generated at eight seconds and used two seconds of contributes two seconds, not eight.
- Track CPFS per project type rather than overall. A talking piece and a fast cut montage have very different profiles, and averaging them together tells you nothing useful.
The value of CPFS is that it prices the thing you sell. Clients buy finished duration. If a client asks for a 60 second video, your CPFS multiplied by 60 is your floor before any of your time is counted, and anything below that is a loss you have chosen not to notice.
The four costs people leave out
Almost every underestimate comes from the same four omissions. Add them back and the number moves a long way for most people, which is the point at which the pricing conversation gets honest.
What gets forgotten and why it matters
| Cost | Why it gets missed | How to count it |
|---|---|---|
| Failed generations | They feel like practice rather than production | Count every attempt on a project against that project |
| Client revisions | They arrive after you have mentally closed the job | Assume a revision round exists and price it in from the start |
| Finishing steps | Upscaling, audio, captions and exports feel like admin | Add them to the project total, since several are billed separately |
| Your hours | There is no invoice, so it never enters the sum | Pick an hourly figure for yourself and multiply it by tracked time |
The fourth row is the one that decides whether this is a business or a hobby. Time is the largest input in almost every project and the only one with no receipt. If you do not put a figure on your own hour, every quote you give will feel fine and every month will feel thin, and you will never be able to explain the gap.
Where the money actually goes in a generation run
Once you start measuring, the spend concentrates in a few predictable places. None of them are mysterious and most of them are controllable at prompt time rather than at spend time.
Cost drivers and what to do about each
| Driver | Why it burns budget | The fix |
|---|---|---|
| Vague prompts | The model has to guess, so more attempts miss | Specify subject, light, lens and motion before generating |
| Complex motion | Hard moves fail more often, so the keeper ratio drops | Simplify the move, or split it into two simpler shots |
| Long durations | You pay for the whole clip and often use a fraction | Generate closer to the length you will actually cut to |
| Generating before storyboarding | Clips get discarded for not fitting, not for being bad | Decide the shot list first, then generate against it |
| Regenerating for taste | Endless variations chasing a preference, not a fault | Set a cap per shot and take the best of that set |
| High settings by default | Every attempt costs more, including the failures | Explore at lower settings, commit at higher ones |
The last row is the single most effective habit on the list. Find the composition and the motion cheaply, then spend properly only on the version you already know works. The storyboarding discipline that stops the fourth row from happening is in [how to storyboard an AI video before you generate clips](/blog/how-to-storyboard-an-ai-video-before-you-generate-clips), and specifying motion so it comes back reliably is covered in [how to control camera motion in AI video](/blog/how-to-control-camera-motion-in-ai-video).
How to measure your keeper ratio in one week
You cannot improve a ratio you have never counted, and counting it takes almost no effort. One week of honest logging gives you a number you can price against for months.
- Open a sheet with four columns: date, shot description, attempts, kept yes or no.
- Log every generation as you make it, including the ones you abandon halfway through a session.
- At the end of the week, divide total attempts by total keepers. That is your K.
- Multiply K by what one generation costs you today to get your cost per usable clip.
- Divide by the average length you actually use from a keeper to get your CPFS.
Then look at the shots with the worst ratios, because that is where the money went. In most logs the failures cluster around a small number of shot types rather than spreading evenly, which means the fix is usually one specific habit rather than general improvement. If your worst rows are all complex motion, you have a motion problem. If they are all people, you have a subject problem. Either way you now know where to aim.
Why the same video costs one creator far more than another
Two people can deliver visually similar 30 second videos and have completely different costs. The difference is almost never the tool. It is a small number of decisions made before anything was generated.
- Whether the shot list existed first. Generating without one produces clips that are individually fine and collectively unusable.
- Whether the prompts were specific. Vague prompts convert directly into attempts, and attempts convert directly into spend.
- Whether hard shots were avoided. Choosing a shot that generates reliably instead of one that looks impressive in your head saves more than any settings change.
- Whether exploration happened at low settings. Paying premium rates to discover you do not like a composition is the most avoidable cost in this medium.
- Whether the same footage was reused. One well built clip that serves three videos has a third of the cost per video.
That last point compounds harder than any of the others, because it changes the denominator rather than the numerator. Getting several finished pieces from one production run is the closest thing to a structural cost reduction available here, and the method is in [repurpose one AI video into a week of content](/blog/repurpose-one-ai-video-into-a-week-of-content).
Turning your cost into a price
Cost sets your floor. It does not set your price. Pricing from cost alone is how skilled people end up earning less as they get faster, because every improvement in your workflow hands the saving to the client rather than to you.
- Calculate the floor: CPFS multiplied by the finished duration, plus your tracked hours at your own rate, plus an allowance for the revision round you know is coming.
- Compare that floor to what the work is worth to the client, which is a function of what it does for them and not of what it cost you.
- Quote the higher of the two by a clear margin, never the floor itself.
- Quote the project, not the components. Clients buy an outcome and do not want a bill of materials for credits.
- Re-measure after every third project. Your CPFS drops as your keeper ratio improves, and that gain belongs to you.
The full method for setting the number rather than just the floor is in [how to price AI video services for clients](/blog/how-to-price-ai-video-services-for-clients) and the pillar guide at [how to price AI video for clients](/guides/how-to-price-ai-video-for-clients). For what the resulting numbers tend to look like as an income rather than a per project fee, see [how much can you earn as an AI video creator](/blog/how-much-can-you-earn-as-an-ai-video-creator).
Bringing the number down without cheapening the work
There are two ways to reduce cost. One reduces waste and the other reduces quality. Only the first is worth doing, and it happens to be the more effective of the two.
- Explore cheap, commit expensive. Settle the composition and the motion at low settings, then spend properly once on the version you know works.
- Storyboard before you generate. Clips discarded for not fitting the edit are pure waste and are entirely preventable.
- Reuse a proven prompt block instead of rewriting from scratch. A library of prompts that have already worked raises your keeper ratio directly.
- Cap attempts per shot. Set a number before you start, take the best of that set, and move on rather than chasing a preference.
- Generate in batches. Setup cost is per session, not per clip, and batched clips also match each other better in the edit.
- Cut the hard shot. If a shot has failed several times, the cheapest remaining option is usually a different shot that says the same thing.
None of those reduce what the viewer sees. They reduce the amount you paid for things the viewer never sees, which is where nearly all the money goes in an unmeasured workflow.
When spending more is the correct decision
Cost control is not the goal, margin is. There are cases where the cheaper route is the more expensive one once the whole job is counted.
- A hero shot that carries the whole piece. One shot usually does most of the persuading, and it deserves the attempts.
- Anything with a face the audience will look at closely. Rescuing a nearly right face in an edit costs more time than regenerating it did.
- A first project for a client who could become a retainer. The economics of that job are about the second job, not the first.
- Portfolio work you will use to win business for a year. Spread across the work it wins, the spend stops being the deciding factor.
- Fixing a structural problem at generation time rather than in post. Grading around bad lighting always costs more hours than regenerating the shot.
The rule underneath these is simple: spend where a viewer or a buyer will notice, save everywhere else. Most cost anxiety in this medium comes from spending evenly across a project rather than concentrating it where it changes the outcome. If you are still deciding which kind of work to build around, the paths are laid out at [start an AI video side hustle](/use-cases/start-an-ai-video-side-hustle).
What to tell a client who asks why it costs anything
This question comes up, usually phrased as an observation that the computer made it. The answer is not defensive and it is not a lecture about credits.
What a client is buying is a finished piece that works, delivered on a date, with someone accountable for it. The generation is the cheapest part of that and always has been. What costs is deciding what the video should say, choosing shots that will actually generate, discarding the attempts that missed, matching them into a sequence, and being the person who fixes it when a shot fails on the last day. None of that is automated, and none of it is what a client wants to do themselves.
Quote a project fee, describe the deliverable and the timeline, and leave the mechanics out of it. Nobody asks a photographer to itemise shutter actuations. The moment your quote reads like a bill for compute, you have invited a negotiation about compute, and that is a negotiation nobody in this business wins.
Short, practical drops on income paths, prompts that work, packaging for views, and getting paid. No spam, unsubscribe anytime.
Frequently asked questions
How much does it cost to make an AI video?
There is no single figure, because the cost is set by your keeper ratio rather than by the advertised price of one generation. Total the credits spent across every attempt on a project, add the finishing steps and your own tracked hours, then divide by the finished duration you kept. Check current credit and plan pricing on Higgsfield directly before you quote anyone, since model costs and tiers change.
What is Cost Per Finished Second?
It is the total project cost divided by the number of seconds that survived into the final export. It matters more than cost per generation because it counts the attempts you discarded, which cost exactly the same as the ones you kept. Since clients buy finished duration, CPFS multiplied by the requested length is your floor before any of your time is added.
Do failed generations really need to be counted?
Yes, and leaving them out is the most common reason a quote loses money. A discarded clip cost the same as a kept one. If you keep one attempt in four, each usable clip cost four generations, and pricing as though it cost one means you are absorbing three out of every four you paid for.
How do I lower the cost of making AI videos?
Reduce waste rather than quality. Explore compositions at low settings and commit at higher ones, storyboard before generating so clips are not discarded for not fitting, reuse prompt blocks that have already worked, cap attempts per shot, generate in batches, and replace any shot that has failed repeatedly with a simpler one that says the same thing. None of those change what the viewer sees.
Should I charge clients based on what the video cost me?
No. Cost sets your floor, not your price. Pricing from cost means every improvement in your workflow hands the saving to the client instead of to you. Work out the floor so you know what to decline, then price against what the finished video is worth to the client and quote a project fee rather than an itemised list of components.
Last reviewed by David on August 21, 2026
