Cost Per Finished Clip: AI Video Economics, Take Ratios, and How to Stop Torching Credits

Alright. This is the one that ties the whole first act together. Over the last several episodes we've been working a single ladder, the single shot. We picked a first tool and learned to read the leaderboard. We took apart the anatomy of a video prompt. We compared image-to-video against text-to-video. We did the episode on aspect ratio, duration, resolution, and frame rate, the constraints episode, the one where I first said draft cheap, finish sharp. We did seeds, negative prompts, and failure modes, where the whole idea was seed discipline. And we did native audio versus silent clips, plus editing by conversation instead of re-rolling. Today is the money episode. This is the discipline that makes all of those primitives actually pay off, because every one of them is really a lever on a single number, and most people never look at that number until their credit balance is gone.

So let me start with the core mindset shift, because everything else in this episode hangs off of it. There are two prices in AI video, and they are not the same price. The first one is the price you see advertised. Call it cost-per-generation. That's the price to press the button one time. Forty cents a second, ten cents a second, whatever the model page says. It looks like the price. It is not the price you pay. The number you actually pay is cost-per-finished-clip. That's the total credits you burn across every roll, every draft, every failed generation, every audio pass, and every upscale, before you have one shot you'll actually ship. Those two numbers start out looking close, and then they diverge fast.

Here's why they diverge. These tools are non-deterministic. The same prompt gives you something different every single time you run it. And because of that, it is genuinely easy to keep regenerating a clip, searching for the perfect version, and burn through twenty dollars in credits without even noticing it happened. That's not me being dramatic, that's straight out of the FluxNote pricing guide for 2026, and it matches what every one of us has actually done at two in the morning chasing a slightly better take. So picture this. You see a forty-cent video. Feels cheap. But you rolled it eight times before you got one you liked. That forty-cent video is a three-dollar-and-twenty-cent finished clip. Same shot. Same prompt. The advertised number was forty cents. The real number, the cost-per-finished-clip, was three twenty. That gap, that multiplier between what's advertised and what you pay, that's the entire subject of this episode.

Now let me define the word finished, because it's doing a lot of work here. A finished clip, or a usable clip, is the one you keep after selects. Selects is just the editing term for choosing the best take out of many. You shoot a bunch, you pick the keepers, those are your selects. In AI filmmaking, the working assumption is that you generate significantly more than your final cut, and you keep only the best clips. How much more? The MindStudio cost breakdown for 2026 says to expect to generate three hundred to five hundred seconds of raw footage for a three-minute short. Do that math per finished minute. A three-minute short is your final cut. Three hundred to five hundred seconds of raw to fill it. That's roughly one hundred to one hundred sixty-six seconds of raw generation for every single finished minute that ships. Which is a two-to-five-times multiplier, on seconds alone, before anything makes it into the edit.

Let me give you the cleaner planning number, because it's easier to carry in your head. The realistic iteration rate is three to five takes per final output. That's the Kling cost analysis figure, and it lines up with the raw-footage math. So your honest planning number, the one you should actually budget against, is the headline per-clip price multiplied by three to five. Not the headline price. The headline price times your take ratio. If you remember nothing else from this episode, remember that one multiplication. Per-second price, times three to five takes. That's your real cost.

Let me ground that with the worked baseline from MindStudio, because it makes the abstraction concrete. They priced out a mixed Runway-plus-Kling short. It came to thirty-four cents per finished second. That's forty-eight dollars spent across one hundred forty clips, totaling five hundred seconds of raw generation, to produce a three-minute film. Read that again. One hundred forty clips generated. Five hundred seconds of raw. For three minutes of finished film. And the all-in number for a full short lands somewhere between sixty-one and one hundred sixty-five dollars. Mid-range, eighty to a hundred thirty. Budget tier, forty to sixty dollars if you use Kling plus free music. Hold onto that range, forty to a hundred sixty-five dollars for a short, because we're going to come back to it at the very end and it proves the whole thesis.

Okay. Let me give you the actual pricing landscape as of June 2026, and I want to put a big asterisk on this up front. These numbers were current to about June 2026, and every one of them is tied to a real model or lab page. But prices in this space move monthly. They churn. So treat what I'm about to say as a snapshot and a set of relationships, not as gospel, and cross-check every number on the live model page before you commit real money. The unit itself even differs by host. Some bill you per second of output. Some bill per video. So know which one you're being charged before you start.

Let's do the labs first, per second of output. Veo 3.1, from Google. Standard tier is forty cents a second at both seven-twenty and ten-eighty p, and sixty cents a second at 4K. The Fast tier drops to ten cents a second at seven-twenty, twelve cents at ten-eighty, thirty cents at 4K. And there's a Lite tier, five cents a second at seven-twenty, eight cents at ten-eighty. Audio is included by default on Veo, which matters, hold that thought. The older Veo 2 is thirty-five cents a second. One wrinkle: some third parties quote Veo 3.1 at seventy-five cents a second when it's routed through Vertex, so the same model can cost you nearly double depending on who's hosting it. Host-dependent. Always check.

Now Sora, from OpenAI. Sora 2 Standard is ten cents a second at seven-twenty. Sora 2 Pro is thirty cents a second at seven-twenty, fifty cents at ten-twenty-four, and seventy cents a second at ten-eighty. There's a batch tier that's roughly half off, but you accept up to twenty-four hours of latency for that discount. Put that in finished terms. A ten-second standard clip is one dollar. A ten-second Pro clip at full ten-eighty is five dollars. Five dollars, for one ten-second roll, at the top tier. Remember that five-dollar number too, it's going to anchor a comparison in a minute. Two housekeeping notes on Sora: the consumer Sora app has been discontinued, and the Sora 2 API has reportedly got a sunset date of September twenty-fourth, 2026. Reportedly. Verify it before you build a pipeline on top of it.

Now the aggregators and hosted models, the cheaper end, per second. Wan, that's 2.5 and 2.6 from Alibaba, runs roughly five to seven cents a second, up to ten-eighty p and fifteen seconds, about thirty-five cents for a five-second clip. Kling 2.5 Turbo Pro is seven cents a second on fal, and Kling 3.0 shows up in some listings as low as roughly three cents a second, but verify that one, it's aggressive. Seedance, from ByteDance, the 1.5 Pro tier gets as low as about two and a half cents a second; the 2.0 Fast tier is around nine cents a second at ten-eighty, roughly eleven cents for a five-second clip, which works out to about ninety videos for ten dollars. Atlas Cloud calls Seedance the cheapest production-quality option out there, and the numbers back that up. Hailuo 02, from MiniMax, on fal: Pro at ten-eighty is eight cents a second, seven-sixty-eight p is four and a half cents, five-twelve p is around one and seven-tenths cents a second, roughly twenty-eight cents per video. And Vidu Q3 is about seven cents a second at ten-eighty p up to sixteen seconds. So that whole tier, Wan, Kling Turbo, Seedance, Hailuo, Vidu, lives in the two-to-nine-cents-a-second range. Keep that band in mind, because that's your drafting territory.

Then there are the credit-based labs, where you're not buying seconds directly, you're buying credits and spending them at a model-specific rate, which hides the real cost a little. Runway prices credits at one cent each in the portal. Gen-4 Turbo is five credits a second, so five cents a second. Gen-4.5 video is twenty-five credits a second, twenty-five cents a second. And if you route Veo 3 through Runway it can run up to forty credits a second. On the consumer side, Runway's free tier gives you a hundred twenty-five credits one time; Standard is twelve dollars a month for six hundred twenty-five credits; Pro is twenty-eight dollars a month for twenty-two hundred fifty; Max is seventy-six dollars a month for ninety-five hundred. Effective per-clip cost lands somewhere between six and twenty-three cents.

Luma's another credit lab. A ten-second Ray3 clip at ten-eighty is about eight hundred credits. Their Standard plan is thirty dollars a month for a hundred twenty generations, Pro is ninety dollars a month for four hundred. And then Kling on the consumer side, which is the best live example of price churn I've got. Standard is about seven dollars a month for six hundred sixty credits. Pro is about twenty-six dollars a month for three thousand. Premier is about sixty-five dollars for eight thousand. And Ultra is now about a hundred eighty dollars a month for twenty-six thousand credits, and here's the kicker, that's up forty-one percent from a hundred twenty-eight dollars in under six months. Forty-one percent, half a year. That is exactly why I keep telling you these numbers churn. On the credit mechanics: a five-second clip is ten credits on Standard versus thirty-five on Professional. A ten-second clip is twenty credits versus seventy. And Kling 2.6 with native audio runs fifty to two hundred credits per video. Oh, and Kling's subscription credits expire monthly, no rollover, which we'll come back to in the gotchas.

Before I leave pricing, let me name the multipliers, because these are the dials that quietly move your cost. Resolution: 4K is roughly one and a half to two times the ten-eighty rate on Veo; going from five-twelve p up to ten-eighty p on Hailuo is a three-to-five-times jump. Duration is linear per second, so a ten-second roll is just twice a five-second roll, no surprise there. Fast versus pro mode is a big one: Kling Standard to Professional is about three and a half times the credits, and Veo Fast to Standard is four times. And native audio: it's bundled free on Veo and Sora, but on Kling 2.6, turning on native audio pushes a clip from ten-to-seventy credits up to fifty-to-two-hundred. So audio, resolution, mode, and duration are four separate knobs, and each one multiplies your real cost-per-finished-clip.

Okay. That's the landscape. Now let's talk about what to do with it, and the first move is the draft-cheap, finish-sharp workflow. This is the constraints episode idea, the one we applied to resolution and duration before, applied now to dollars. The principle is simple. You lock the two things that are hard to fix later, composition and motion, on the cheapest possible path. Composition is your framing and layout, where everything sits in the frame. Motion is your camera move plus your subject action. You nail those down cheap. And only once the shot is approved do you spend on the expensive render.

So what are your draft levers? Lowest resolution, five-twelve or seven-twenty p. Fast mode, not pro. The shortest duration that still shows the motion, which is often five seconds. And the cheapest model that can still read your composition. That's your draft stack. Then you reserve ten-eighty p, 4K, pro mode, longer durations, and the premium model for the approved hero shot, and only that shot. Let me make the savings concrete with that five-dollar number from earlier. One ten-second Sora 2 Pro render at ten-eighty is five dollars. That same five dollars buys you about fifty seconds of Veo 3.1 Lite drafts at five cents a second. Fifty seconds. That's ten separate five-second draft rolls before you've spent the five dollars on the one hero render. Ten cheap looks at your composition for the price of one premium clip.

It works the same in the credit world. A Kling five-second Professional roll is thirty-five credits. The Standard draft of that same five seconds is ten credits. So you can draft three and a half times for the price of one pro roll. Draft on Standard, lock the shot, then spend the Professional credits once. And here's the cheapest draft trick of all: drafting stills is cheaper than drafting video. Gemini's Nano Banana line reportedly gives you about twenty free images a day in the app, and around fifty requests a day up to 2K resolution with no watermark in Google AI Studio. So you mint your keyframe for free, in an image model, pick the one good frame, and feed that single frame into image-to-video, instead of rolling text-to-video blind and praying. That ties straight back to the mint-keyframes episode and the image-to-video versus text-to-video episode. The cheapest pixel is the one you generated in an image model.

Now, the part nobody tells you about until it's already cost you money. The hidden-cost gotchas. There are seven of these and they all bite. First: you pay for completed generations even when they're bad. Runway's own refund policy says it plainly: if a generation completes, even if the result doesn't match your prompt, the credits are consumed. And on platforms like Meshy, the credits are deducted the moment you press generate. So a roll that completes and looks nothing like what you asked for still costs you full price. The button doesn't care if you're happy.

Second gotcha: failed and errored generations are inconsistently refunded. Users across Veo, Kling, and Sora report credits deducted without ever getting a video back. There are Adobe community threads and Google Flow threads full of people saying the gen errored out and the credits were gone anyway, sometimes four times in a row. So the safe way to budget is to assume failed rolls cost money. Don't plan as if errors are free retries. Some of them are. Plenty of them aren't.

Third, and this one is subtle because it feels free: small conversational edits are full regenerations. When you say make the jacket red, the tool doesn't reach in and recolor the jacket. It re-renders the entire clip from scratch and bills you full credits for it. Now, editing by conversation is genuinely useful, we did a whole episode on it, it saves you prompting effort and it preserves your intent across turns. But it does not save you credits. Every conversational turn is a brand-new paid generation. So love it for the workflow, don't trust it for the budget.

Fourth: audio passes re-render the whole clip on platforms where audio is a separate step. We saw the number already. On Kling, adding native audio swings a clip from ten-to-seventy credits up to fifty-to-two-hundred. So that's not a small add-on, it's a multiplier on the whole clip. Fifth: upscale and extend are billed separately, on top of the original. Upscale is re-rendering at a higher resolution, extend is generating extra seconds, and both are full additional charges. And upscale failures have eaten credits too, there's a Firefly report of generations failing but still getting charged. So your finished cost isn't the render, it's the render plus the upscale plus the extend, each billed on its own.

Sixth gotcha, and this one quietly wrecks your seed discipline: the enhance-prompt or rewrite toggle. When that toggle is on, the tool rewrites your prompt behind the scenes, which changes the output, and sometimes changes the cost. Worse, when enhance-prompt is on, the same seed no longer reproduces the same result. So the seed discipline we worked so hard on in the seeds episode just stops working, and you end up re-rolling to try to get back something you already had. Turn it off when you care about reproducibility. And seventh: subscription credits expire at the end of the month, no rollover, Kling says so explicitly. Pay-per-use packs avoid the expiry, but they cost more per unit. On Kling, the Pro pack is about one and two-tenths cents per credit versus the Standard pack at about one and a half cents per credit, so you trade a higher per-unit price for not losing them at month-end. Pick based on whether you actually burn your monthly allotment.

Alright. Let me hand you the actual copyable budgeting workflow, six steps, because knowing the gotchas only helps if you've got a procedure. Step one: compute your cost-per-finished-second up front. Take the per-second price and multiply it by your real take ratio, and start that ratio at three to five. So Veo 3.1 Standard at forty cents a second, times four takes, times a five-second shot, is eight dollars for one finished five-second shot. Not two dollars. Eight. If you budgeted two, you're already three-quarters short before you start.

Step two: set a per-shot roll budget and a stop rule, and write the number down. MindStudio's guidance is a maximum of three regenerations per clip before you move on. And the line I want you to tape to your monitor: a slightly imperfect clip that fits the edit is better than a depleted credit balance. So three rolls, and then you either select what you've got or you change the approach, a new prompt, a new seed, a new reference image. What you do not do is roll number four of the exact same thing hoping the dice land better.

Step three: batch your drafts. Generate your candidate variations in one cheap pass, then select from them. Do not go roll, judge, roll, judge, roll, one at a time, because that serial loop is exactly where the spiral starts. Fire off a batch, look at them together, pick. Step four: reuse seeds to avoid re-exploring ground you've already covered, that's the seed-discipline episode, and turn off enhance-prompt so the seed actually holds, which is gotcha number six from a minute ago. Step five: tier your models. Use the cheapest model that clears the bar for drafts and transitions, Seedance, Hailuo, Wan, Veo Lite, that whole two-to-nine-cent tier. Save the premium model for hero shots and emotional beats only. And step six: use the leaderboard's value columns. The Artificial Analysis Video Arena ranks models by blind-vote Elo, real people voting on which clip is better without knowing the source, and it lists price right alongside, specifically dollars per minute of video for a ten-eighty p, five-second, twenty-four-frame clip. So you can sort by quality and by price together. But, and this matters, bench your own shots. The leaderboard winner overall isn't always the best model for your specific motion or your specific subject. Run your actual shot through a couple of contenders and judge for yourself.

Now I want to name the pitfall directly, because it has a name and recognizing it is half the battle. It's the regeneration death spiral. Sunk-cost rerolling. It's chasing a marginally better roll, just one more, one more, and torching your budget on diminishing returns. It is the sunk-cost fallacy with a generate button attached. You've already spent the credits, so you keep spending to justify the credits you spent. Here's how to recognize you're in it. You're past your roll budget. The improvements are shrinking with every roll, each one barely better than the last. And you're re-rolling the same prompt, hoping the dice land better, instead of changing your approach. If those three things are true, you're in the spiral, stop.

How do you break it? Four moves. First, lock a good-enough select. The moment you have a roll that would work, save it as a real selected take, formally, before you roll again. Now every additional roll has to beat a banked option, not just beat nothing. That changes the psychology completely. Second, branch, don't replace. Keep your current best, and explore a variation alongside it. Never overwrite your select chasing a maybe, because the maybe is often worse and then you've lost the good one. Third, fix small things downstream. A slightly-off color, a cut that's a touch too fast, a minor blemish, that belongs in post, not in a new generation. You color grade, which is just adjusting color and contrast for the look you want, and you trim in the NLE, the non-linear editor, your timeline software, Premiere or DaVinci or CapCut, whatever you use. That's zero generation cost. You're paying with your time, not your credits. And fourth, cap it numerically. Three rolls, then you select or you branch. The spiral, fundamentally, is just the absence of that cap. The cap is the cure.

Let me widen the lens to cost at scale, because the discipline gets more important the more you produce, not less. Variations multiply everything. Say you need the same product shot in five aspect ratios across three backgrounds. That's fifteen finished clips. Now apply the three-to-five-times take multiplier. That's forty-five to seventy-five paid rolls for one product, one shoot. At that scale, the choice of the cheapest acceptable draft model and a hard roll cap doesn't just save you a little, it's the difference between a sane invoice and a blown budget. The multiplier compounds across every variation.

And the right budget depends entirely on what the clip is for. A one-off social clip tolerates a budget-tier model and seven-twenty p. You finish it on a five-to-ten-cent-a-second draft model, no upscale, and the whole thing comes in at single dollars. Don't gold-plate a throwaway. A client deliverable is a different animal. That justifies premium hero renders, ten-eighty or 4K, native audio, an upscale pass. But, and this is the whole point, the discipline does not change. You still draft cheap to lock the composition, and then you spend the five-dollar Sora Pro render, or the forty-cent-a-second Veo render, once, on the approved shot. Same workflow, higher ceiling.

And that brings us right back to the number I told you to hold onto. A full short still lands somewhere between forty and a hundred sixty-five dollars. That range is the proof of the entire thesis of this episode. The money doesn't live in the headline price. A forty-cent-a-second model and a five-cent-a-second model aren't eight times apart in your final invoice, because the final invoice is dominated by your take ratio, how many times you rolled, not by the sticker on a single roll. The money lives in the take ratio. Control the take ratio and you control the budget. That's it. That's the discipline.

Let me close with the verification note, because I'd be doing you a disservice if I let you treat any of these numbers as permanent. Every figure I gave you was current to about June 2026, and every one was tied to a real model or lab page. But prices move monthly. Kling Ultra jumped forty-one percent in under six months. The Sora 2 API has a reported September 2026 sunset. So re-check the live model page before you commit, and send yourself, and your shots, to the Artificial Analysis Video Arena to bench your own work, because the right model for your motion and your subject is something only your own eyes can confirm. That closes out Act One, the single shot. We've climbed from typing a prompt and taking what we get, all the way to landing the shot we pictured on a budget we control. Next act, we start stitching shots together into consistent scenes. But you don't get to a finished film if you torched your credits on shot one. Draft cheap, finish sharp, cap your rolls, and spend the premium render once. That's the money discipline. See you next episode.