Cost per Usable Clip: The Real Economics of AI Video

TL;DR: AI video vendors don't price a result, they price an attempt. Some bill per second, some per generation, some per frame, so raw sticker prices compare different things. Divide the price by how often an attempt is actually usable, add the vendor's separate upscale charge, and the real ai video cost per clip runs several times higher than the number you compared.

What Does "Cost per Usable Clip" Actually Mean?

Every AI video vendor publishes a price. Almost none of them publish the number that matters: how much you'll actually spend to get one clip you keep.

A generation and a usable clip are not the same unit. A generation is one API call, billed the moment it runs, whether the output has a warped hand, a subject that drifts and morphs halfway through, sits there with almost no motion at all, or a camera move that ignores what you asked for. None of that voids the charge. The vendor delivered a video. Whether it's the video you can cut into something is your problem, not theirs.

Cost per usable clip is the sticker price divided by your keep rate: the share of generations you actually put in a timeline rather than delete. If you kept every attempt, cost per usable clip and cost per generation would be the same number, and a vendor's pricing page would be all the math anyone needs. Nobody's keep rate is 1.0. The gap between the two numbers is the real cost of the workflow, and it's a number a pricing page can never show you, because the vendor has no way to know your subject, your prompt quality, or your own bar for "usable."

Why Doesn't One Price-per-Second Number Work Across Vendors?

Because vendors don't all bill the same axis. Line up sticker prices from four vendors' own pricing pages and you're comparing per-second rates, a flat per-generation rate, and a rate that only applies after generation is already finished.

Luma bills Ray 3.2 by the generation, not the second. Its own API pricing page describes the model in two words: "Pay per video." 1080p is priced at $1.20 for a 5-second clip and $3.60 for 10 seconds (read September 3, 2026). That's a 3x jump for 2x the duration, which only makes sense if the price is set per output tier, not multiplied by elapsed time. Veo 3.1, Wan 2.7 and Vidu Q3 all bill per second instead, at rates that scale roughly linearly with duration. Runway's Gen-4.5 generation is priced per second too, in credits: 12 credits, or $0.12, per second of generated video, confirmed on Runway's own help center article on how credits work (read September 3, 2026). But Runway's separate video upscaler bills per output frame, not per second of the source clip, which means Runway's own two numbers aren't on the same axis as each other, let alone as everyone else's.

None of this is hidden. It's just easy to miss when you're skimming five pricing pages in five tabs, mentally converting everything to "dollars for a ten-second clip," and losing the actual billing unit along the way. Here's what four vendors bill, in their own words.

How Much Do the Major Vendors Actually Charge?

The table below comes from each vendor's own current pricing or API reference, fetched and read on September 3, 2026. One warning before you read it: resolution tier names are not comparable across vendors. MiniMax's top tier is labelled "2K" and never "1080p." BytePlus's 4K tier exists only on an older Seedance version, not the current one. Read each vendor's own tier names rather than assuming a shared resolution ladder underneath them.

A few of these need context the table can't carry.

Seedance 2.5's own number is a promo, not a list price. BytePlus's own documentation states: "From 14:00 (UTC+8) on August 14, 2026 through 14:00 (UTC+8) on September 17, 2026," 1080p output is billed at "28% off the list price" with "prices starting at approximately USD 0.41 per second" (read September 3, 2026). There is no published list price behind that discount, and 480p and 720p output aren't eligible for it at all. Quoting $0.41/second as Seedance's ongoing rate after September 17 would be quoting a number the vendor never actually committed to.

MiniMax runs a genuinely separate re-render path, and it's cheaper than generating twice. MiniMax-H3-Regeneration takes an existing 768P output and re-renders it to 2K for $0.05 a second of the regenerated output, per MiniMax's own pricing docs (read September 3, 2026). That's cheaper than generating fresh at the native 2K rate of $0.13 a second, which is a real argument for generating cheap first and upgrading only the clips you actually keep.

Legacy MiniMax models still bill flat per video, the same way Luma does. MiniMax-Hailuo-2.3-Fast is $0.19 for a 6-second 768P clip and $0.32 for 10 seconds, not a linear per-second rate, and a different billing shape from MiniMax's current H3 line on the same pricing page (read September 3, 2026).

How Do You Actually Calculate Cost per Usable Clip?

The formula is simple. The hard part is being honest about the denominator.

cost_per_usable_clip = (price_per_generation × attempts_per_keeper)
                      + (upscale_price, if the keeper needs one)

Illustration only, not a measured rate: say you're generating 10-second clips at $0.15 a second, so $1.50 a generation, and hypothetically you keep one attempt in four. Your real cost per usable clip is $6, not $1.50. Now say that keeper needs a 4K pass. Runway's own pricing guide works a 10-second, 30fps example directly: 210 credits ($2.10) at 720p/1k, 270 credits ($2.70) at 2k, 360 credits ($3.60) at 4K, on top of whatever the original generation already cost (read September 3, 2026). Add that to the $6 and the clip that looked like $1.50 on the pricing page is closer to $9.60.

Two things move that number more than any vendor's per-second rate ever will.

Your keep rate, not the vendor's price. A vendor that's twice as expensive per second but produces a keeper on the first attempt three times as often is cheaper in the number that actually matters. This is exactly why "cheapest per second" round-ups are close to useless on their own: they answer a question nobody is actually asking.

What you count as usable. A stiff pan you can crop around in the edit is a keeper. A subject that drifts off-model mid-clip usually isn't, no matter how you edit it afterward. Our companion piece on what actually gets fixed in post versus what needs a re-roll walks through exactly where that line sits. Tightening your own definition of "usable" before a shoot starts is the cheapest lever in this entire equation, and it's free.

Track your own keep rate for even a week: attempts made, clips kept, no more math than that. Do it before trusting any number in this post, including the illustration above. It will differ by subject, by vendor, and by how tight the prompt was going in.

What Happens if You Prepay a Year for a Model That's Being Retired?

This is the sharpest version of "sticker price isn't the real price," and it isn't hypothetical.

OpenAI discontinued the consumer Sora web and app experiences on April 26, 2026, per OpenAI's own help center (read September 3, 2026). The Sora API, meaning sora-2, sora-2-pro, and every dated snapshot, is a separate surface, and it's scheduled to shut down September 24, 2026: three weeks from the date of this post. OpenAI's own deprecations page states it verbatim: "On March 24th, 2026, we notified developers using the Videos API and Sora 2 video generation model aliases and snapshots of their deprecation and removal from the API on September 24, 2026." Its recommended-replacement column for every Sora row is blank (read September 3, 2026). There is no named successor.

If you're paying an annual plan to any tool built on that API, September 24 is your real expiration date, whatever your renewal date says. A yearly seat bought in July doesn't buy twelve months of Sora access. It buys whatever's left between purchase and September 24, then a forced migration to a different model with a different prompt syntax, a different resolution ladder, and, per the numbers above, a different price entirely.

That's cost per usable clip taken to its logical end. The sticker price on a subscription page assumed a year of access that the underlying model was never actually going to provide.

What Should You Check Before You Commit Budget to a Vendor?

A short list, in the order it actually saves money:

1. What's the billing unit? Per second, per generation, or per frame. Read the vendor's own pricing page for this; a round-up that lists "$0.10/second" for a vendor that bills flat per video has already lost the axis.
2. Does the price include the resolution you need, or is that a separate call? If it's separate, get the upscale endpoint's own price before you budget, not after.
3. Is the number a current list price, or a time-boxed promo? Seedance's own 0.41/second figure above is a live example of the second case.
4. Are reference images, input video, or directed audio billed too? Several vendors bill reference material on top of the output, at a fraction of the output rate, and audio isn't always free.
5. What's your own keep rate on this vendor, measured, not guessed? This one number moves the real cost more than any per-second rate comparison.
6. If a deprecation notice exists, what's your actual expiration date? Not the renewal date on your invoice. The date on the vendor's own deprecations page.