The hidden line item: what failed renders cost

Every vendor quotes a price per generation. Production runs on a surplus and then selects, so the honest unit is attempts divided by keeps. Here is the arithmetic, and a calculator for your own numbers.

Buying creative/By Badal Kariwal/01 Sep 2026/10 min read

What is in here

What is a realistic cost per AI video ad?
The surplus is the product, not the waste
Why one in two and one in ten are different businesses
The three costs that never reach the invoice
How do you raise a keep rate?

The short answer

Take the price of one generation and divide it by the share of generations you would actually run. That is the real unit, and it usually lands two to five times above the number on the invoice. A ten-dollar clip at a one-in-two keep rate costs twenty dollars. At one in ten it costs a hundred. The money is the smaller half of the bill: reviewing, re-briefing and the shot you eventually discover you cannot get at all are paid in hours nobody quoted.

What you get out of this

The one line of arithmetic that turns a price per generation into a price per usable clip
A calculator that runs your own price, your own keep rate and your own monthly volume
Why a one-in-two shop and a one-in-ten shop are running different businesses, not the same one at different quality levels
The three costs that never reach an invoice, and which of them is worst
What actually raises a keep rate, and what only looks like it does

The four numbers this whole page turns on

Two ours, two theirs

43%of films that clear every automated check we run are still killed on the second watchSutra Haus production log

2xthe minimum surplus we build before selecting - ten statics made so five can be sentSutra Haus

5%of new creatives become winners, and the rate is the same at top accounts as at average onesMotion Creative Benchmarks 2026

5.99winners a month against 1.75 at the average, on the same spend, from 31 new creatives a week instead of 11Motion Creative Benchmarks 2026

The first two are ours, counted in this studio's own production log. The second two are Motion's, from 550,000 ads across more than 6,000 advertisers and roughly $1.3bn of spend. Ours are a keep rate. Theirs is a win rate. They are different failures and they stack.

01What is a realistic cost per AI video ad?

Somewhere between two and ten times the price of one generation. There is no single figure, because the figure is a division: what one attempt costs, divided by the share of attempts you would put in front of a customer. Every vendor publishes the numerator. You supply the denominator, usually without ever having written it down.

Say a clip costs eight dollars to make and you keep two out of every five. Your real cost is twenty dollars a clip, and on top of that you have spent whatever your own hour is worth watching the three that died. Nobody sells you a keep rate, because nobody can sell you one. It is a property of your brief, your product and your taste, not of the model.

What a keep rate does to a $10 generation

Same clip, five keep rates

See the numbers as a table

Keep rate	Cost per usable clip
9 kept in 10	$11
1 in 2	$20
1 in 3	$30
1 in 5	$50
1 in 10	$100

Arithmetic, not a measurement: ten dollars divided by the keep rate, rounded. The thing worth noticing is that the bottom rows are not a worse version of the top rows. They are a different budget with a different bottleneck.

The rate also moves. Week one on a new brand is the worst week you will have, because you are still learning what the product looks like when it is lit badly, and what its owner will not accept. We build a surplus for that reason rather than as a flourish. If you want this arithmetic run across routes instead of inside one, what AI ads, UGC creators and an agency each cost does the three-way version.

Nobody buys renders. You buy the ones you would run, and you pay for the rest of them anyway.
The sentence we open a scoping call with

Your price, your keep rate, your volume

Put your own numbers in

Price of one generation

Share of generations you would actually run

Usable ads you need each month

Minutes you spend judging one generation

Cost per usable ad-

Generations you have to buy-

Generation spend a month-

Hours a month spent judging-

Four inputs, because four is all it takes. The review figure is the one people leave blank and it is usually the largest number on the page: a generation you reject still costs you the minutes it took to decide. Nothing here includes a subscription, an editor, or the cost of running the ad.

02The surplus is the product, not the waste

Selection is where quality comes from. That reads like a slogan until you price it. Build one thing and correct it, and every correction lands on the most expensive object you own. Build ten and choose five, and the last creative act of the project is a choice between finished things rather than a repair on an unfinished one.

The volume argument is not ours and it is stronger than anything we could measure alone. Motion looked at more than 550,000 ads and found the winner rate is about 5% at good accounts and about 5% at ordinary ones. What separates them is swings: the best large accounts ship 31 new creatives a week instead of 11, and land 5.99 winners a month against 1.75 on the same spend.

Correct one thing, or choose between ten

Two ways to spend one budget

Correct one thing, or choose between ten
Dimension	Build one, fix it	Build a surplus, select
Total generation spend	Yes. Lower	No. Two to five times higher
Where the last decision happens	No. On a finished render	Yes. On a shortlist of finished things
What a change costs	No. Unbounded, because nobody agrees when it is done	Yes. Nothing. You pick a different one
What you are actually choosing between	This version and a slightly different version of it	Genuinely different swings at the same brief
Who has to reject things	Nobody, which is why nothing gets rejected	Somebody, named, before the client sees anything
What a bad week looks like	No. One ad, five rounds, still wrong	Partly. A thin shortlist, and you can see it early

We run the right-hand column and we lose the first row on it. Building a surplus genuinely costs more in generation than building one careful thing. It costs less in everything else, which is the trade and it is not free.

There is a limit to the argument and it is worth stating. Craft has a measurable effect on cost, and it is small: CreativeX, across roughly 822,000 observations, put a ten-point rise in their creative quality score at about a two percent fall in CPM. Motion measured a different thing on a different scale: winner count against weekly cadence. We would not put a multiplier between the two, and neither should anyone selling you that. Craft is a floor you have to clear, not the lever you pull.

Which side of the table the surplus sits on

A studio that builds ten and invoices for the five it sent has absorbed the failure rate. A studio that builds ten, sends ten and asks you which you like has handed it back to you, at your hourly rate rather than its own. Both are legitimate businesses. Only one of them is the one you thought you were buying.

03Why one in two and one in ten are different businesses

At a one-in-two keep rate the bill is money and it is small. At one in ten the bill is hours and it is yours. Twelve usable ads at one in ten means a hundred and twenty generations to sit through. At four minutes each, watched properly with sound, that is eight hours of somebody's month spent saying no.

The crossover between the two regimes is not a fixed rate, because it depends on what an hour of your attention is worth against what a generation costs. Put both into the calculator above and the point where the hours column overtakes the spend column is your own crossover. For most people running a cheap model and their own eyes, it arrives around one in five.

The same pipeline at two keep rates

One panel each

The bottleneck is generation spend

You are buying about twice what you ship. Review is a normal afternoon, the shortlist is real, and the arithmetic is boring in a good way. The number to watch is the invoice.

The failure mode here is quiet. A high keep rate can mean the brief is well understood, and it can equally mean every variant is the same variant with a new grade. Five versions of one concept is one swing rendered five times, and it will fatigue as one.

Cost per usable ad is about twice the sticker price
Review time is manageable, so nobody measures it
Real risk: a set that looks varied and is not
Test for it: describe two of the five to a stranger and see if they can tell them apart

What you pay in: Money
Watch: Sameness across the set

Neither column is the good one. A one-in-ten keep rate is normal on a new brand in week one, and a one-in-two rate on a mature brand usually means the set has stopped taking risks. What matters is knowing which one you are in, because the fix is different.

04The three costs that never reach the invoice

Generation spend is the part everybody counts because it arrives as a number in an email. The other three are real money and they arrive as time, mood and a hole in the shot list.

Where the rest of it goes

One at a time

Cost 1

The reviewing

Every generation you reject still costs the minutes it took to decide. Four minutes sounds like nothing until you multiply it by the attempts, and the attempts are set by the keep rate rather than by the shipping plan.

There is a second-order cost inside it. Judgment degrades across a session. The hundredth clip gets a worse look than the tenth, and the errors it lets through are the expensive kind, because a defect that survives review is now in a finished ad.

Cost 2

The re-briefing

A rejected batch has to be explained before it can be replaced, and the explaining is skilled work. Notes like make it better and try again are not briefs, they are hope, and they produce another batch with the same fault.

A re-brief that changes one variable is worth ten that change everything. Ten changed at once tells you nothing about which change did the work, and that is how a project ends up on round five with no more knowledge than it had on round one.

Cost 3

The shot you cannot get at all

Some shots do not come. Not badly, not yet: at all. A hand doing something delicate with the product, a specific material behaving the way it behaves, a room you have no photograph of. You can spend an entire budget discovering this one attempt at a time.

Our rule is that a missing shot gets named out loud and escalated the moment it is suspected, never silently swapped for something weaker that fits the slot. The silent swap is what produces an ad where every beat is fine and the film is dead.

1 / 3

None of these three are on a pricing page anywhere, including ours. The third one is the expensive one, and it is the only one that gets worse the longer you leave it.

The rule that saves the most moneyWhen a shot will not come, we name the gap and say exactly what would make it available - a photograph from a particular angle, a twenty-minute phone shoot, a different beat entirely. A refusal that names the frame that would make it legal turns a dead end into a short shopping list. A refusal that quietly substitutes something weaker turns it into a finished ad nobody can defend.

A woman in a white t-shirt and wide beige trousers walking through a hard shaft of sunlight past a dark stone colonnade — One beat of a finished film. The cost of this shot is not the cost of this shot - it is the cost of this one plus the ones that came back with the light wrong, the walk wrong, and a shadow the building could not have cast.

The fourth cost is the week you spend running the same batch again with slightly different words, because nobody wrote down what was wrong with the first one in terms specific enough to act on. That week costs more than the generations did, and it is entirely avoidable by writing one honest sentence about why a take died.

05How do you raise a keep rate?

By moving decisions earlier, not by generating better. Almost every kill we log traces to something decided on paper: the wrong story, an undefined hook, a shot list assembled from what happened to be on disk. Model quality sets the floor of what a keep rate can be. The brief sets the ceiling, and the ceiling is lower.

Six things that move a keep rate, in order of how much they move it

Tick as you go - it remembers

Decide the story on paper first. Write the film as sentences before anything renders. If the sentences can be shuffled without becoming false, you have an assortment rather than a story, and no render will fix that.Name the hook, in the brief, as a mechanism. Not a mood, not a look. What physically happens on screen in the first half second, and why a stranger's eye goes to it.Cheap mockup before the full build. Styleframes and a motion test of only the riskiest beat. Minutes of compute against hours, and it kills the bad ones while they are still cheap.Change one variable per re-brief. Ten changes at once teaches you nothing. One change teaches you which change worked, and the knowledge compounds across the month.Ask for the pace, not the duration. A seven-second request for a two-second action gets you slow motion, not spare footage. Say what happens and how fast, then let the length follow.Have somebody who did not make it watch it cold. At phone size, with sound, among other content. The maker is the worst judge of the maker's work, and the gap is measurable rather than theoretical.

The first three are free and cost minutes. The last three cost real time. We run all six, and we still kill 57 of every 100 films that pass every automated check we own, which is the honest ceiling on what process can do for taste.

Four finished statics from one brand

Four survivors

All four ran. They are the visible half of a set built roughly twice this size, which is why they can be this different from each other - a set built at one to one has to accept its first idea. The ones that did not survive cost about what these cost.

The number you should ask any partner for is not their price. It is how many things they made and did not send you. A studio that shows you everything it built has no selection layer, and you have just been hired as one. How you tell the two apart across a whole month rather than a single batch is the subject of telling a real winner from noise, and how many swings you need before the question is even answerable sits in how many creatives you actually need.

Four things buyers say about generation cost

Flip them

All four were said to us by people with a budget open in another tab. Two of them are half right, and the half that is right is the half about the invoice.

Questions people actually ask

Open what you need

What is a realistic cost per AI video ad?

Two to ten times whatever one generation costs, set entirely by your keep rate. An eight-dollar clip kept two times in five is twenty dollars a usable clip. The same clip kept one time in ten is eighty. On a subscription, divide the monthly fee by the ads you actually ran rather than by the credits the plan gave you. Any number quoted without a keep rate attached is a price per attempt, and attempts are not what you run.

What is a good keep rate for AI video generation?

We do not have an industry benchmark to give you and we have not found one with a stated sample, so treat anyone quoting a figure with suspicion. Our own working forecast is one usable film in two or three, and the honest measured number from inside our pipeline is that only 43% of films clearing every automated check survive a human watching them.

How much does AI ad production actually cost per month?

Generation spend is usually the smallest line. Take your monthly volume, divide it by your keep rate to get attempts, multiply by the price per attempt, then add the review hours at whatever your time is worth. For most people the review hours overtake the generation spend somewhere around a one-in-five keep rate.

Is there a published AI render failure rate?

Not one we would cite. Model vendors publish neither, and the number is not really a property of the model anyway - the same model at the same settings produces very different keep rates on a well-briefed job and a vague one. The nearest published thing is winner rate in market, which Motion puts at about 5% of new creatives across 550,000 ads.

Should I pay per generation or per month?

Per generation punishes a low keep rate directly, so every rejection shows up on the invoice. A subscription hides it, and the cost reappears as your own hours and a credit allowance that runs out mid-project. Neither is cheaper by nature. Pick the one whose failure mode you would notice, because the one you cannot see is the one that grows.

How do I work out my own keep rate?

Look at last month. Count everything that was delivered to you, then count what actually went live. That fraction is your keep rate, and it is almost always lower than the number people quote in meetings. Do it per brand rather than in aggregate, because a single difficult product can drag a whole account's average somewhere misleading.

Every price in this market is quoted per attempt because attempts are the only thing a seller controls. The number that decides your year is the one neither of you has written down yet, and you can have it by Friday: count what arrived, count what ran, divide.

Where the numbers came from

Motion. Creative Benchmarks 2026: winners are rare - 550,000+ ads, 6,000+ advertisers, about $1.3bn in spend; used for the winner rate and the 5.99 against 1.75 comparison
CreativeX. Creative Quality Score - roughly 822,000 observations; used for the size of the craft effect against the size of the volume effect

Every figure above links to the place it was published. Numbers marked as ours are measured inside this studio and we say so where they appear. We do not print a statistic we cannot point at.

Badal Kariwal

Runs Sutra Haus, a one-person ad studio that has shipped over a thousand finished creatives - film and stills - for DTC brands and hotels. Writes here about what the work actually taught him, including the parts that failed. The person who reads your brief is the person who builds the work. Send him something to make.

One attempt, priced at zero

Send a link to your product. Get one finished ad back.

No call, no deck, no invoice. One finished cut built from your own product inside three days, yours to run whether or not we ever work together. It is one attempt on our side of the ledger, which is the only honest way to show you a keep rate.

Replies within a day. Ad within three.