AI Video Pricing Comparison: What 10 Seconds Really Costs

Ten seconds of finished video costs $0.50 on one AI video model and $10.20 on another — a 20× spread across models sold for the same job. We priced 13 AI video models in one unit on September 2, 2026, reading every rate off the vendor's own page. The number that moves your real bill most is one none of those pages publishes.

Cheapest 10-second clip: $0.50 (Runway Gen-4 Turbo). Dearest: $10.20 (Seedance 2.5, 1080p). Audio doubles Veo 3.1 from $0.20 to $0.40 per second. Google's page says you are charged only for successful videos and Kling's says failed credits are refunded; OpenAI's and Runway's pricing pages say nothing either way — and where a page is silent, retries are a live billing risk. Across 714 logged generations on our own stack only 55–73% succeeded, so plan on 1.37× to 1.80× the attempts for one usable clip.

What one 10-second clip costs on 13 AI video models

Every vendor quotes a different unit: Runway prices in credits worth a cent each, Google and OpenAI per second, fal some models per video, and Google bills audio as its own line. So any AI video pricing comparison starts with arithmetic — convert everything to one finished ten-second clip, and carry each rate's published resolution with it rather than pretend the tiers match.

Three rows need a caveat before comparing like for like:

• Runway's own Gen-4 rates and fal's Kling listing name no resolution on the rate card.
• Sora 2's $0.10 is a 720p rate; its 1080p tier is Sora 2 Pro at $0.70.
• Hailuo 3's $0.15 is quoted at 2K, above the rest of the table.

The spread is not a quality ladder

Gen-4 Turbo at $0.50 and Seedance 2.5 at $10.20 are not twenty rungs apart on anything a viewer would name. An AI video model's price tracks compute, resolution ceiling and how new the weights are — not how the output lands.

The same AI video model shows up at several counters

Six of the thirteen rows are not sold by the lab that trained them: Veo 3.1, Seedance, Wan 3, Hailuo 3 and Grok Imagine all appear on Runway's rate card as hosted third-party models, and fal sells Kling and Veo 3 beside its own catalogue. An AI video model's price is set by the counter you buy it at — so price the same model twice before committing a project to whichever dashboard you opened first.

Audio doubles Veo 3.1, and most AI video tables hide it

Google publishes Veo 3.1 twice: $0.40 per second with audio, $0.20 without. Same model, same resolution, exactly double for the soundtrack — which moves a ten-second clip from $2.00 to $4.00, a larger swing than changing vendor. Veo 3.1 Fast repeats the shape: $0.15 with audio, $0.10 without.

A table quoting a single Veo number has to pick one of those rates, and rarely says which. Budget from the audio-inclusive rate for a project you will score in post and you plan at twice the real cost; budget from the silent rate when you needed dialogue and your first invoice is a surprise.

The rule: read an AI video rate card for its switches, not just its model names. On Google's page resolution and audio are separate multipliers; on Runway's, resolution is. Where both apply, they compound.

Which AI video vendors charge you for a failed generation?

Here is the question that rarely appears in an AI video comparison table and is worth more than the model choice: when a generation errors out, or a content filter refuses it, do you pay for it? Four pricing pages, read the same day:

Google and Kling both commit in writing, which is worth saying plainly.

A written policy and a refunded balance are not the same event

A Kling user on the Adobe community forum reported burning 375 credits twice — 750 total — on 15-second 1080p generations that never produced a video, with the credits still deducted. The expert reply is the honest version of how this works: credits "are deducted when you press generate. If the generation fails, they should be restored, but it is not necessarily instantaneous." No confirmation was posted that they came back.

So it is a two-way split with a wrinkle: Google and Kling commit in writing, OpenAI and Runway leave you to find out, and a written commitment says the balance should return — not when. For a studio running hundreds of AI video generations a week on a prepaid balance, that timing is a line item worth confirming before the invoice.

The AI video retry tax: what retries actually cost you

Every price above assumes the generation works. None of them do, all of the time — and none of the five vendor pages we read publishes a success rate.

We can publish ours. Over the 30 days to September 2, 2026 our stack logged 714 generations that reached a verdict — a finished file or a returned failure:

At 68% you press generate roughly three times for every two clips kept. That multiplier is real for everyone; what it costs depends on the policy in the section above — which is why the two questions belong on one page:

• Where the vendor charges only on success (Google) or refunds (Kling), retries cost time and queue position, not money. A $4.00 Veo 3.1 clip still bills $4.00; you wait about half an attempt longer for it.
• Where the pricing page is silent (Sora, Runway), the rate card cannot tell you whether a refused or errored AI video generation is billed. If it is, that 68% turns $4.00 into $5.87 — a bigger swing than most model-versus-model gaps, and unresolvable from the published price.

So the AI video retry tax is not a surcharge you can look up: zero on two of these vendors, unknown on the other two, multiplied by a success rate none of them disclose.

Two caveats, because they change how you should use our numbers:

• These are our models and our users, not an industry benchmark. Read 55–73% as one measured stack, not as your number.
• Most of our failures are content-policy refusals, not crashes. A refusal is a different animal from an outage — and on a vendor whose page never mentions moderated generations, it is exactly the case you would want covered in writing.

What we would not do is take a list price at face value: your own success rate multiplies every number here, and it is the largest uncosted item in an AI video budget.

Measure your own AI video success rate first

The fastest way to find your success rate is to run one prompt several times and count how many outputs you would ship. Kavel's AI video generator opens on a free engine — the cheapest place we know of to run that count before committing a budget to any rate card.

Our per-credit rates for every AI video model, in the same one-finished-second unit, are in AI video cost per second: what 16 models actually charge; plan pricing is on the pricing page.

What the cheapest AI video option actually is

If the deliverable is ten finished seconds and nothing else is specified, Runway Gen-4 Turbo at $0.50 is the cheapest rate on this list, with Kling 2.5 Turbo Pro at $0.70 next and Veo 3.1 Lite at $0.80 the cheapest that names 1080p on its own rate card.

But "cheapest per second" and "cheapest per usable clip" are different questions, and only the second is the one you are asking:

• An AI video model at $0.50 that lands one clip in two takes two attempts.
• One at $0.80 that lands four in five takes barely more than one.

If the vendor bills failures, the first is also the dearer of the two. Until vendors publish success rates, that is a comparison you run yourself, on your own prompts.

FAQ

How much does a 10-second AI video cost in 2026?

Between $0.50 and $10.20 across the 13 AI video models we priced. Runway Gen-4 Turbo is cheapest at $0.05 per second; Seedance 2.5 dearest at $1.02 for 1080p. Most land between $1.00 and $4.00 for ten seconds. Budget attempts as well as seconds — at 70% success you press generate about 1.4 times per clip kept.

Why is Veo 3.1 listed at two different prices?

Because audio is billed separately. Google publishes Veo 3.1 at $0.40 per second with audio and $0.20 without, so a silent ten-second clip is $2.00 and the same clip with sound is $4.00. Veo 3.1 Fast is $0.15 and $0.10. Runway's hosted Veo rows carry the same split, so a lone "Veo 3.1" AI video figure is always ambiguous.

Do AI video generators charge you for failed generations?

It depends on the vendor, and only some tell you. Google's Veo page states you are charged only if the video is generated successfully, and Kling's credits policy states failed generations are refunded. OpenAI's and Runway's pricing pages do not address it at all. One published user report describes credits deducted and not visibly returned, so check your balance rather than assume the policy settles instantly.

What is the cheapest way to make 1080p AI video?

Runway Gen-4 Turbo ($0.05/second) and Kling 2.5 Turbo Pro ($0.07/second) are the lowest AI video rates we priced, though neither names a resolution; Veo 3.1 Lite at $0.08 is the cheapest publishing 1080p. In practice the cheapest route is whichever model lands your prompt first time — on our stack that took 1.37 to 1.80 attempts per output.

Is Sora 2 more expensive than Veo 3.1?

At the top tier, yes: Sora 2 Pro is $0.70 per second at 1080p versus $0.40 for Veo 3.1 with audio. At the entry tier it reverses — Sora 2 at 720p is $0.10, above Veo 3.1 Lite's $0.08 at 1080p. OpenAI's batch rate is half price, bringing Sora 2 to $0.05.

Sources

All AI video rates read on September 2, 2026 from each vendor's own page:

• Runway API pricing — Gen-4 Turbo, Gen-4.5, and the third-party models Runway hosts
• Google Gemini API pricing — Veo 3.1 Standard / Fast / Lite, audio and silent rates, charge-on-success statement
• OpenAI API pricing — Sora 2 and Sora 2 Pro standard and batch rates
• fal pricing — Kling 2.5 Turbo Pro, Wan 2.5, Veo 3
• Kling credits policy — refund-on-failure wording

Success rates are our own, measured across 714 completed generations on kavel.ai between August 3 and September 2, 2026.

Last updated: September 2026