What AI video generation costs per second

The whole field in one unit

Every rate here is per second of generated output, converted to a finished minute by multiplying by 60. Two conversions are ours and are stated so you can check them: Kling publishes credits, and its credits policy prices top-ups at "$1 USD = 66 Credits", making a credit $0.015152; Runway publishes credits and its developer pricing page states credits are "$0.01 per credit". Everything else is already in dollars.

Seedance and the Runway-hosted Veo rows are priced through Runway's developer pricing at $0.01 a credit. Gemini Omni Flash is billed on tokens, not seconds: Google prices video output at $17.50 per million tokens and states a rate of 5,792 tokens per second of 720p, which is where the ~$0.101 comes from. Rows marked "gone 24 Sep" are recorded for the historical record — see the Sora rate card. One caveat on the Seedance 2.5 rows, added 23 September: Runway prices that model on two meters, not one, listing "20 credits per second of output, plus 10 credits per second of input" at 480p and 30/15 and 68/34 at 720p and 1080p. The rates tabulated above are the output halves, which is the whole bill for a text prompt; feed it video and the true rate is 50% higher — $0.30, $0.45 and $1.02 a second — so a video-to-video job at 1080p costs $61.20 a finished minute rather than $40.80. No other model on this card publishes an input rate.

The floor is a commodity, and three products sell it

The cheapest second of generated video is $0.05, and it is not a single vendor's loss-leader. Google sells it as Veo 3.1 Lite at 720p with audio included. Runway sells it twice, as Gen-4 Turbo and as Act-Two, both at 5 credits a second. Those three are not substitutes for each other — Act-Two is a performance-capture tool, and Runway publishes no audio rate for either of its two — but the price has clearly settled, and at $3.00 a finished minute the generation layer is not where a small project's money goes.

The ceiling is 30x higher. Seedance 2.0 at 4K is 150 credits a second through Runway, which is $1.50 a second and $90.00 a finished minute — and the same model at 1080p is 40 credits, so Seedance charges +275% for the step from 1080p to 4K. Kling charges the same premium for its own 4K step, 8 credits to 30, also +275% — though that one is not quite like-for-like, because Kling's 8-credit 1080p rate is explicitly silent and its documentation does not state whether the 30-credit 4K rate includes native audio. That two vendors setting prices independently landed on a similar 4K multiplier is worth noting rather than a rule, but it does suggest where the field thinks 4K sits.

Sora had a cheap tier after all, and it was on a different page

Correction, 23 September 2026. This site's Sora review published on 15 September described Sora as having "no cheap tier at all, its $0.10 minimum being exactly double Veo 3.1 Lite's $0.05". That is true of the standard column only. OpenAI's API pricing page also carries a batch column for both video models, at exactly 50% of the standard rate on all four configurations — and CartSignal had not published it.

So on batch, Sora's floor was $0.05 — level with Veo 3.1 Lite, not double it, and the model everyone spent this month leaving held a joint-cheapest second in the field. It changes the migration arithmetic too: this site's exit table measured every move off Sora as cheaper or break-even, which holds, but measured from the batch rates those margins roughly halve, and from the batch floor the best available exit is a tie rather than a saving. The documented worst case also halves: a 20-second sora-2-pro generation at 1080p was $14.00 standard and $7.00 on batch, and the 120-second extended maximum was $84.00 and $42.00.

Why it was missed is the more useful part, and it extends a pattern this site has now recorded four times on the same product. The batch column appears on OpenAI's pricing page. It does not appear on the sora-2 or sora-2-pro model reference pages, which were re-read today and show a single standard column — the same two pages that, one day from removal, still carry no deprecation notice of any kind. The batch route itself is documented in the video generation guide, which says "Use the Batch API when you need to queue many video renders for offline processing, review pipelines, or studio workflows" and that "Batch-generated videos are available for download for up to 24 hours after the batch completes". Three surfaces, three different amounts of the story.

Does Google match it? Not on any page we can find, and the honest answer is that this is unresolved rather than settled. Google's Batch API documentation states that "Batch API usage is priced at 50% of the standard interactive API cost for the equivalent model", but it lists no models itself, saying only to "Refer to the Models page for each model's support of Batch API" — and the Gemini API pricing page carries no batch column for any Veo row. CartSignal therefore does not publish a discounted Veo rate and does not claim Veo is excluded. If you are moving a high-volume, non-interactive pipeline off Sora batch, whether Veo is batch-eligible is worth an hour of your own testing, because it is worth half your bill.

Resolution and duration are coupled, and only one vendor says so plainly

This is the finding most likely to change what you actually spend, and it is not a price at all. Google's Veo documentation gives durationSeconds allowed values of "4", "6", "8" — and then states the duration must be "8" when using 1080p or 4k resolutions, reference images, or the extension feature. You cannot buy a four-second 1080p Veo clip. Asking for the higher resolution silently doubles the shortest thing you are allowed to order.

For anyone iterating on short shots this matters more than the per-second rate. A team testing forty variations of a four-second cutaway pays $64.00 at 720p on Veo 3.1 Standard; the same forty variations at 1080p are $128.00 — exactly double, because each one is now eight seconds whether the extra four are used or not. The per-second price did not change at all — it is $0.40 either way.

Published minimums exist elsewhere too, and they bite hardest on the cheap models. Runway's developer card gives Aleph 2.0 a 56-credit minimum at 28 credits a second, a two-second floor; and Seedance 2.0 Mini a 64-credit minimum at 16 credits a second, which is a four-second floor on the cheapest Seedance variant. A one-second Mini clip bills as four.

Nobody sells you a minute

Every per-minute figure on this page, including the ones in the headline table, is a multiplication rather than something you can order. The longest single generation any of these models documents is twenty seconds, and that one is being withdrawn tomorrow.

The asymmetry in that last column is the capability the Sora shutdown actually removes, and it is not a price. Google allows extension by "up to 7 seconds per extension" across "up to 20 extensions" for a combined "up to 148 seconds of video" — but the documentation states the extension feature is limited to 720p. OpenAI documented extension to a 120-second total with 1080p available on sora-2-pro. So a long continuous 1080p piece was buildable on Sora and, on the extension route, is not buildable on Veo. A 148-second Veo 3.1 Lite piece at 720p costs $7.40; the 1080p equivalent has to be cut together from separate eight-second generations with no continuity guarantee at all.

This is also why this page declines to rank the models on anything but price. Two models at the same rate can be nothing alike on screen, and a stitched minute from a model with poor shot-to-shot consistency is worth less than a stitched minute from a good one at twice the price. For where each tool fits by job rather than by rate, see video tools by job and Veo vs Kling.

Audio is priced three incompatible ways

The same second of video, with and without a soundtrack, is priced by three different logics — and two of them are applied to the same model depending on who sells it to you.

• Included, no discount for dropping it. Google states audio is included by default on all three Veo 3.1 variants and publishes no silent rate. OpenAI included audio at all four Sora rates.
• A 50% surcharge. Kling charges a flat half again for native audio at both resolutions — 6 to 9 credits a second at 720p, 8 to 12 at 1080p — plus "2 Credits/s" for voice control on top.
• A 50% discount for dropping it. Runway, reselling Veo 3.1, publishes 40 credits a second with audio and 20 without; on Veo 3.1 Fast it is 15 and 10, a 33.3% saving.

The practical consequence is narrow but real. If you are laying a scripted voiceover or a music bed over the result — which most commercial work does — Google bills you $0.40 a second for generated audio you will mute, while Runway will sell you the identical model silent for $0.20. That is the one configuration where the reseller beats the vendor outright, and it exists because Runway chose to publish a rate Google does not.

A credit is not a credit: the same second at four prices

Runway is the clearest case on the site of a single product with several meters. Its developer portal sells credits at $0.01; its subscription plans sell the same credits at whatever the plan price divided by the allowance works out to. Gen-4.5 costs 12 credits a second on every one of them.

A 3.0x spread on an identical second, decided entirely by which door you came through. Unusually, the top subscription tier beats the developer list price — Max annual at $0.008 is 20% under it — which is the reverse of what this site found in voice, where developer APIs undercut every studio subscription by several times. It only holds if the whole allowance is spent, and on Standard and Pro it expires monthly.

The allowance is the harder limit anyway. At 12 credits a second, Runway's 625-credit Standard month buys 52 seconds of Gen-4.5 — under a minute of finished video for the whole month — Pro buys 3.1 minutes, Max buys 13.2 minutes, and the free plan's 125 one-time credits buy about ten seconds, once. In this lane the plan allowance binds long before the per-second price does, which is exactly the pattern this site found in AI dubbing, where a single localised video consumed an entire month.

What one real job costs

A 30-second product clip at 1080p, priced across the field, with a column for the five takes a real shoot burns before something is usable.

Two honest adjustments to those numbers, both upward. On Veo at 1080p the clip is four eight-second generations, so you buy 32 seconds rather than 30 — $2.56 rather than $2.40 on Lite, and $12.80 rather than $12.00 on Standard. On Kling, 30 seconds exceeds the 15-second ceiling, so it is two generations and a join. Neither changes the ordering, and both illustrate the page's main point: the iteration count and the clip ceiling move this bill more than the rate card does. Going from one take to five multiplies every row by exactly 5; choosing Lite over Standard divides it by 5. Those two dials are the same size, and only one of them is on a price list.

What this page does not publish

• Any quality or prompt-adherence claim. No video was generated for this page. Two models at the same rate may be nothing like each other on screen, and in this lane that gap is probably larger than the price gap.
• A Veo batch rate. Google documents a 50% Batch API discount and does not state whether Veo is eligible; the pricing page shows no batch column for video. Unresolved, not excluded.
• Which resolutions Runway Gen-4.5 supports. Runway's developer pricing lists Gen-4.5 as a single flat 12 credits a second with no resolution rows, unlike Seedance which is broken out by resolution. A flat rate is not evidence that every resolution is available at it, so no 4K claim is made here in either direction.
• Seedance bought direct. ByteDance sells Seedance on token-based pricing through BytePlus ModelArk, which cannot be converted to a per-second rate without an assumed tokens-per-second figure. The Seedance rows here are Runway's resale rates, which are published per second. See the Seedance profile.
• Which Kling plans may generate 4K. Kling publishes the 30-credits-a-second 4K rate on its credit cost guide with no plan restriction, while its feature page carries a gating row; the membership page renders its prices client-side. This has been unresolved on the Kling profile since 16 September and remains so.
• Third-party tracker rates. Several roundups currently quote Veo 3.1 at $0.03 a second. That figure appears nowhere on Google's pricing page, which was read today and gives $0.05 as the cheapest Veo second. Rates here come from vendor pages only.

What to buy

For drafts, social cutdowns and anything at 720p: Veo 3.1 Lite at $0.05 a second with audio included is the cheapest complete product in the field, and the 148-second extension route works at that resolution. For 1080p commercial work: Veo 3.1 Fast at $0.12 is the value rung — Standard costs 3.3x for the same 1080p frame and charges the identical $0.40 at 720p, so on Standard the lower resolution is simply worse value for the same money. If you will replace the audio anyway: buy Veo 3.1 silent through Runway at $0.20 rather than $0.40 direct. If you need a long continuous piece: check the extension limits before the rate card, because at 1080p none of these will give you one. And whatever you pick, budget the takes, not the seconds — five re-rolls is a 5x multiplier on every figure on this page, which is larger than the gap between most of the vendors on it.

Frequently asked questions

How much does it cost to generate one minute of AI video?

Between $3.00 and $90.00 at published rates, before re-rolls. The cheapest finished minute is $3.00 — Veo 3.1 Lite at 720p with audio, or Runway Gen-4 Turbo and Act-Two at 5 credits a second. Veo 3.1 Fast at 1080p is $7.20, Kling at 1080p with audio is $10.91, Veo 3.1 Standard is $24.00, Kling 4K is $27.27, Veo Standard 4K is $36.00, and Seedance 2.0 at 4K through Runway is $90.00. Five takes of a 30-second clip moves the bill further than most vendor choices do.

Which AI video generator is cheapest per second?

Three products tie at $0.05: Veo 3.1 Lite at 720p with audio included, and Runway Gen-4 Turbo and Act-Two at 5 credits on the $0.01 developer credit. They are not substitutes — Act-Two is performance capture, and Runway publishes no audio rate for either. Until 24 September 2026, sora-2 on the batch tier tied that floor at $0.05.

Is audio included in AI video generation pricing?

Three incompatible policies. Google includes audio at every Veo rate with no silent discount, and OpenAI included it at every Sora rate. Kling charges a flat +50% for native audio, plus 2 credits a second for voice control. Runway, reselling Veo 3.1, publishes 40 credits with audio and 20 without — so a silent Veo second is $0.20 through Runway and unavailable at any price direct from Google.

Why can I not order a one-minute AI video?

Because nothing here generates a minute in one pass, so every per-minute figure is arithmetic. Veo 3.1 documents durations of 4, 6 and 8 seconds, so a minute is eight generations and seven joins; Kling documents 3 to 15 seconds; Sora documented 20. Extension routes have their own limits — Google allows up to 148 seconds but only at 720p, while OpenAI documented 120 seconds with 1080p available. No published rate guarantees continuity across a join.

Is it cheaper to buy a video model direct or through Runway?

Veo 3.1 Standard is an exact tie at $0.40 a second. Veo 3.1 Fast is cheaper direct, $0.10 at 720p against $0.15 through Runway. But Runway sells a silent Veo 3.1 at $0.20 that Google does not list at all. The bigger spread is inside Runway: a credit is $0.01 in the developer portal and $0.024 on a Standard monthly plan, so the same Gen-4.5 second is $0.12 or $0.288 — 2.4x — and Max annual at $0.008 undercuts the developer rate.

Sources

• Gemini API pricing: all seven Veo 3.1 per-second rates by resolution, audio included by default, Gemini Omni Flash video tokens, no batch column for video — read 23 September 2026
• Veo documentation: durationSeconds allowed values, the 8-second rule for 1080p and 4K, and the 720p-only 148-second extension limit — read 23 September 2026
• Gemini Batch API: the 50% pricing statement and the absence of a model list — read 23 September 2026
• OpenAI API pricing: sora-2 and sora-2-pro standard and batch columns — read 23 September 2026
• OpenAI video generation guide: Batch API usage, 24-hour batch retention and one-hour download URLs — read 23 September 2026
• Kling VIDEO 3.0 model user guide: 6, 8, 9 and 12 credits a second, the 2-credit voice control add-on and the 3 to 15 second duration range — read 23 September 2026
• Kling credit cost guide: native 4K at 30 credits a second — read 23 September 2026
• Kling credits policy: the $1 USD = 66 Credits top-up rate used for every Kling dollar figure here
• Runway developer pricing: credits per second for Gen-4.5, Gen-4 Turbo, Act-Two, Aleph 2.0, the Veo 3.1 audio and silent rows, all Seedance variants, published minimums and the $0.01 credit price — read 23 September 2026
• Runway plan pricing: Free, Standard, Pro and Max prices and monthly credit allowances — read 23 September 2026

Related pages

• Best AI video tools — the three lanes priced in the same unit, with editing and avatar video alongside generation
• Veo vs Kling — the two cards matched resolution for resolution
• Sora — the final rate card, recorded the week it was withdrawn
• AI dubbing cost per minute — the same treatment for localisation
• Commercial use rights — whether you may publish what you just paid to generate