AI Video Cost Per Second: Why the Sticker Rate Lies

Short answer: the advertised ai video cost per second on a vendor's pricing page describes one clean, accepted generation, not what you actually spend to get a clip you can publish. Your real cost per usable second depends on how many attempts it takes to get a shot that works, whether the resolution or duration you need sits on a higher tier, and whether a rejected or low-quality generation still burns the credits it consumed. Every one of those variables is different for your project than for the demo reel on the vendor's homepage, which is exactly why no article, including this one, can print a single honest number that applies to you. What follows is the arithmetic to do that for yourself, on any tool, before you commit a month's budget to one.

Why the advertised rate is not your rate

Every AI video vendor prices its base unit the same rough way: a rate per second of generated output, sold through a credit system rather than a raw cash meter. Runway's own credits page states plainly that "each model uses a set number of credits per generation depending on the model, duration, and resolution" (Runway help center, checked August 2026), and every other major vendor's pricing page follows the same shape. That single sentence is doing more work than it looks like. It means the sticker rate is not one number, it is a formula with at least three inputs, and the pricing page only shows you the cheapest combination of them.

Compare that to how a print shop prices paper: the per-sheet rate you see in the ad is for the smallest size and the plainest stock. Order a larger sheet or a heavier weight and the unit price moves, even though the ad never changes. AI video credits work the same way, and the headline rate is almost always quoting the smallest, lowest-resolution, shortest-duration option.

The four things that move your real cost

Resolution. Most vendors meter their highest resolution output at a materially higher credit rate than their standard tier, and several gate their top resolution behind a paid plan entirely. Before you compute anything, check which resolution you actually need to publish at, not which one the demo reel uses.

Duration. As covered in why your video comes out in fixed, short chunks instead of one continuous clip, a single generation is capped at a short menu of durations, and longer options inside that menu typically cost proportionally more credits, not the same flat rate stretched further.

Native audio. Where a model offers audio generated alongside the video rather than added afterward, that option is usually billed as a separate, higher tier. If your workflow needs synced sound in the same pass, price that mode specifically, not the silent baseline.

Failed and rejected generations. This is the variable the pricing page never mentions and the one that changes the arithmetic the most. Runway's own credit documentation does not describe any refund or credit-back process for a generation that comes out wrong, and creator reports collected across several 2026 pricing write-ups consistently describe credits being consumed on outputs that were unusable, regenerated, or simply discarded. Treat that as the default assumption for every vendor unless its own terms say otherwise: a generation that fails, glitches, or misses the brief still counts against your balance.

The math the pricing page will not do for you

Here is the calculation, and it works on any vendor's current numbers, which is deliberate: this niche's rates move often enough that any figure printed here today would likely be wrong by the time you read it.

1. Open the vendor's own current pricing or credits page and note the credit cost for the resolution, duration and audio setting you actually need, not the cheapest listed option.
2. Estimate your attempts-per-usable-clip ratio honestly. If you have not tracked this yet, start a simple tally for your next ten generations: how many did you keep versus discard or regenerate. Most creators are surprised by this number the first time they count it.
3. Multiply the per-generation credit cost by that ratio. That product, not the sticker rate, is your real cost per usable clip.
4. Divide by the clip's length in seconds to get your actual ai video cost per second, and compare that to the vendor's advertised rate to see the real gap.

A tool that looks expensive per generation but produces a usable result on the first or second attempt can end up cheaper per usable second than a tool with a lower sticker rate that regularly needs five attempts to land a shot. The sticker rate tells you nothing about that ratio, and the ratio is the part of the cost you actually control through better prompting and shot planning, covered in getting the same character into more than one shot without a face change forcing a re-roll.

Why "cost per second" cannot be compared across tools directly

Even once you have your own number for one tool, comparing it straight across to a competitor is misleading, because vendors do not all meter the same base unit. Some sell a flat per-second rate. Others sell prepaid credit bundles priced in bulk, where the effective per-second rate depends on which bundle size you bought and whether you used it before it expired. A few price the entire generation as a single package regardless of exact duration, so "per second" is not even how they bill internally, it is a number derived after the fact for comparison purposes only. Before comparing two tools, confirm you are looking at the same billing unit for both, on their own current page, not a third-party recap of it.

Where the real waste usually hides

The single biggest lever most creators never adjust is the retry habit itself. Regenerating a clip with the same prompt and hoping for a better roll of the dice is the most expensive way to fix a bad output, because it changes nothing about what caused the failure. Diagnosing the actual problem first, whether it is a hands-and-faces issue, a camera-move misread, or a resolution setting that was never worth paying for on that particular shot, cuts the attempts-per-usable-clip ratio far more reliably than switching vendors does. That ratio, not the advertised rate, is where your monthly spend is actually decided.

Audio generation follows a parallel structure worth knowing if your project needs narration alongside the visuals: it is billed by character or by minute of finished output rather than by second of video, and the same "meter versus sticker price" gap shows up there too, covered in how AI voice pricing actually works, credit by credit.

FAQ

Is a lower advertised cost per second always the cheaper option? Not necessarily. A lower rate on a model that needs more attempts to produce a usable clip can cost more overall than a higher rate on a model that gets it right on the first or second try. Track your own attempts-per-usable-clip ratio before deciding on price alone.

Do failed AI video generations get refunded? Generally treat them as non-refundable unless a vendor's own terms explicitly say otherwise. Checked against Runway's own credits documentation in August 2026, no refund or credit-back process for a failed or unsatisfactory generation is described there, and this pattern is widely reported by creators across other vendors as well.

Does resolution really change the price that much? Yes, on every major vendor's own pricing structure. Higher resolution and native audio generation are typically metered at a separate, higher credit rate or restricted to paid tiers entirely. Check the specific setting you need against the vendor's current page rather than assuming the advertised baseline rate applies.

Why won't this article just tell me which tool is cheapest? Because the honest answer depends on your attempts-per-usable-clip ratio, which is different for every creator and every kind of shot, and because credit rates in this category move monthly. A number printed today would be stale within weeks. The method above works on whichever tool you are pricing, on the day you are pricing it.

How many attempts does it typically take to get one usable AI video clip? There is no honest single answer to this without measuring your own generations, because it depends heavily on the model, the complexity of the shot, and how tightly the prompt is written. Start tracking your own ratio over your next batch of generations rather than relying on a figure from someone else's workflow.