Blog · 25 October 2026 · The KitCut team

AI video generator pricing (October 2026): cost per second, model by model

At list price, a second of AI video from the best-known models costs $0.03 to $0.57 at 720p and 1080p in October 2026. A 10-second 1080p clip with sound costs $0.60 on LTX-2.3 Fast, $1.68 on Kling 3.0, $4.00 on Veo 3.1 and $5.69 on Seedance 2.5. Below is every price we found, each linked to the maker's own page.

Updated 25 October 2026: every price checked that day.

Jump to a model: Veo 3.1 · Omni Flash · Kling · Seedance · Runway · Luma · Wan · MiniMax · Pika · Grok · the cheapest · FAQ

At a glance

• Per second, at 720p and 1080p: from $0.03 (LTX-2.3 Fast, Grok Imagine Video 1.5 Lite, or Veo 3.1 Lite without sound, at 720p) to $0.569 (Seedance 2.5 at 1080p).
• Sound and resolution can double the price, or more. Veo 3.1 costs $0.40 a second with sound and $0.20 without. Luma Ray3.2 costs four times as much at 1080p as at 720p.
• Sora 2 is gone, and Veo 3.1 has left the Gemini API. Google itself now sells Veo 3.1 only on Google Cloud, and on the Gemini API recommends Gemini Omni Flash 1.1 instead ($0.152 a second at 1080p).
• An app plan can cost less than the API, if you use every credit. Kling's plans work out at $0.086 to $0.16 a second of Kling 3.0 at 1080p with sound, against $0.168 on its API; Runway's reach its API's $0.12 only on Max ($95 a month).

AI video generator pricing comparison

The major video models with a public API, from each maker's own page on 25 October 2026:

How to read it: the last price column is one 10-second clip, with sound where the model makes it. Clip and token prices (Seedance, Luma, Hailuo and Gemini Omni Flash) are divided down to one second, and each model's section shows the arithmetic. A model that stops short of 10 seconds needs two clips, billed by the second either way. Luma and Hailuo charge more per second for a 10-second clip, so their column shows that clip's own price. Kling 4.0 is fal's price, as Kling does not sell it on its own API yet. These are API prices: Runway's and Kling's app plans are in their sections, and other makers' app plans are not covered.

What a 10-second AI video clip costs

At 1080p, with the model's own sound where it makes any, one 10-second clip from a model in our main table costs from $0.60 on LTX-2.3 Fast to $5.69 on Seedance 2.5, more than nine times as much. That is one take. Every retake costs the same again.

What changed this autumn: Sora 2 shut down, Veo 3.1 left the Gemini API

Sora 2 no longer has a price. OpenAI's deprecations page records that the Videos API, sora-2, sora-2-pro and their dated snapshots were removed six months after developers were told, with no replacement named, and OpenAI's help article says "The Sora web and app experiences were discontinued on April 26, 2026." Replicate and WaveSpeed still had Sora 2 pages on 25 October. Those listings are stale, and WaveSpeed's warns that generations "may fall back to alternative models without notice". On Google's side, a Gemini API key no longer reaches Veo 3.1: you need Google Cloud (where Vertex AI is now called Gemini Enterprise Agent Platform) or a reseller. The dates that matter for prices:

Veo 3.1 price per second (and Veo 3 pricing)

Google sells Veo 3.1 in three tiers, priced per second of video, with sound optional. From Google Cloud's price page:

A Veo clip is 4, 6 or 8 seconds long, so a 10-second shot is two clips (6 + 4). The full model costs the same at 720p as at 1080p, and turning sound off halves it.

Searching for Veo 3 pricing? Google retired Veo 3 and Veo 3 Fast on 30 June 2026, on Google Cloud and on the Gemini API, and its model versions page names Veo 3.1 as the upgrade. Google Cloud's price page still lists Veo 3 at Veo 3.1's prices at 720p and 1080p ($0.40 a second with sound, $0.20 without), but you can no longer call it there.

Where to buy it: fal sells Veo 3.1 at Google's price but is deprecating it on 17 November, and charges 25 to 50% more for Veo 3.1 Fast with sound ($0.15 a second). Replicate, WaveSpeed, OpenRouter and Runway's API list it too. Google quotes 11 seconds to 6 minutes per generation; the median on Replicate is about 90 seconds.

Gemini Omni Flash pricing

Gemini Omni Flash 1.1, generally available since 27 August 2026, is billed in tokens at $17.50 per million: 5,792 tokens per second of 720p video and 8,688 per second of 1080p. That is $0.101 a second at 720p and $0.152 at 1080p, and sound is always included. The Gemini API price page gives the 720p rate; Google Cloud's gives both. A clip runs 3 to 10 seconds and can be extended to 40.

Google's deprecations page names it the recommended replacement for the Veo 3.1 previews, and at 1080p with sound it costs 38% of Veo 3.1's price. fal, WaveSpeed, Runware and Runway's API sell it too; Replicate charges 48% more. It is first on Arena's text-to-video leaderboard (the board dated 21 September).

Kling AI pricing: plans, Kling 3.0 and Kling 4.0

Kling, Kuaishou's video model (Kuaishou's announcement), is sold two ways by Kling itself: app plans paid in credits, and an API paid in units.

Kling AI plans

From Kling's plan page on 25 October, in US dollars:

If every credit is used, a plan buys Kling 3.0 at 1080p with sound for $0.086 to $0.16 a second, less than the $0.168 that Kling's API charges at list price. The plan page does not say what a second costs in credits, so the seconds use the rate in Kling's Video 3.0 guide, published in February: 12 credits a second at 1080p with sound (9 at 720p, and 8 and 6 without sound). Monthly credits last 31 days, then expire. Premier also comes with 8,000, 28,000 or 38,000 credits a month, and the free Basic plan gets some free credits each month (the page does not say how many).

Kling API pricing

The API bills in units listed at $0.14 each. Kling 3.0 costs $0.084 a second at 720p without sound and $0.126 with it; at 1080p, $0.112 and $0.168; at 4K, $0.42 either way (Kling's developer pricing). Kling 3.0 Omni, without a reference video, costs the same as Kling 3.0 without sound, and $0.112 and $0.14 with it. Kling 3.0 Turbo always makes sound, at $0.112 and $0.14. Clips run 3 to 15 seconds, so a 10-second 1080p clip with sound costs $1.68 on Kling 3.0, and a minute $10.08.

You buy the units in advance: trial packs of 100 units for $14 and 1,000 for $140, valid for 30 days (30% off a first purchase), and 180-day packs from $700 for 5,000 units, with 10% off from 30,000 units. fal sells Kling 3.0 at the same per-second price, and Replicate at twice the price.

Kling 4.0, announced on 28 September, is not on Kling's own API yet: Kling shows preview prices of 1.0 unit a second at 720p and 1.5 at 1080p ($0.14 and $0.21), and says the official pricing will be announced separately. fal sells it now at $0.157 a second at 720p and $0.235 at 1080p, for clips of 3 to 30 seconds; its page does not say whether it makes sound.

Seedance 2.0 price per second, and Seedance 2.5

ByteDance sells Seedance through BytePlus ModelArk, which bills in tokens and states a price per 5-second clip (BytePlus pricing):

• Seedance 2.0: five seconds of 720p are 108,000 tokens at $7.00 per million, which is $0.756, or $0.151 a second. Five seconds of 1080p are 243,000 tokens at $7.70 per million: $1.871, or $0.374 a second. Seedance 2.0 Fast ($0.605 per 5 seconds) and 2.0 Mini ($0.378) stop at 720p. BytePlus's own examples round these to $0.76, $1.87, $0.60 and $0.38.
• Seedance 2.5: five seconds of 720p are 108,000 tokens at $10.70 per million, which is $1.156, or $0.231 a second. Five seconds of 1080p are 243,000 tokens at $11.70 per million: $2.843, or $0.569 a second, the highest per-second price in our table.

Sound is optional and costs the same. Seedance 2.5 makes clips up to 30 seconds and 2.0 up to 15, and a 10-second clip at 1080p costs $3.74 on 2.0 and $5.69 on 2.5.

Two practical limits. BytePlus says its model services are "not available in the United States", and an account costs $30 to activate and runs three jobs at a time for an individual. fal sells a US-hosted Seedance 2.5 at about twice BytePlus's price, and WaveSpeed charges 56% more.

Runway pricing: plans and Gen-4.5 per second

Runway plans

From Runway's pricing page on 25 October, in US dollars. In the app, as on the API, Gen-4.5 uses 12 credits a second:

Runway's API sells Gen-4.5 at $0.12 a second, so a plan matches it only on Max at $95 a month, and beats it only on Max billed yearly ($0.096); on Standard a second costs about two to two and a half times as much. Plan credits also pay for the app's other models at their own rates. On Standard and Pro they reset each month and do not roll over; Max rolls up to a month of credits over.

Runway API pricing

Runway's API charges 12 credits a second for Gen-4.5, at $0.01 a credit: $0.12 a second (Runway's API pricing). It makes 720p-class video only, 2 to 10 seconds long and without sound, so a 10-second clip costs $1.20 and a minute $7.20.

Two things to know before a big batch. Runway charges for generations its moderation blocks, blocks public figures by default and suspends accounts with too many blocked attempts. And a new account (tier 1) runs one or two jobs at a time, and 50 to 200 generations a day per model. Runway's API also resells other makers' models, including Veo 3.1, Gemini Omni Flash, Kling 3.0, MiniMax H3 and Seedance.

Luma Ray3.2 pricing

Luma prices Ray3.2 per clip, without sound (Luma's API pricing). A 5-second clip costs $0.30 at 720p, or $0.06 a second, and $1.20 at 1080p, or $0.24 a second. A 10-second clip costs three times the 5-second one, not twice: $0.90 at 720p and $3.60 at 1080p, or $0.09 and $0.36 a second. Two 5-second clips cost $0.60 and $2.40, if the shot can take a cut in the middle. Replicate, fal and WaveSpeed sell it too.

Wan 3.0 pricing

Alibaba's Model Studio lists Wan 3.0 at $0.10 a second at 720p and $0.20 at 1080p, with sound and clips up to 30 seconds (Alibaba Cloud pricing). Those are the prices of its international deployment, in Singapore; its global deployment, in the US, Germany and Japan, lists about $0.083 and $0.165. On 25 October both carried a "limited-time 30% off" label. Wan 3.0 Prime costs $0.14 and $0.28, and Wan 2.7 costs $0.10 and $0.15, with clips up to 15 seconds.

Wan 3.0 leads Artificial Analysis's text-to-video leaderboard with audio, but it is not quick: Alibaba quotes 1 to 5 minutes per clip. The newest Wan with open weights is Wan 2.2, under the Apache-2.0 licence (GitHub): you can run it yourself, or pay Model Studio $0.10 a second for it at 1080p, in 5-second clips without sound.

MiniMax H3 and Hailuo pricing

On MiniMax's own API, H3 costs $0.08 a second at 768p, with sound. It has no 1080p tier; 2K costs $0.13 a second. Clips run 4 to 15 seconds, so a 10-second 768p clip costs $0.80 and a minute $4.80. The same page lists two cheaper 768p tiers, H3-Highspeed at $0.06 a second and H3-Lite at $0.04. H3 Max, which fal post-trained on H3, costs $0.08 there; fal also sells it, at a promotional price that ends on 31 October. The older Hailuo 2.3 is sold per clip: $0.28 for 6 seconds or $0.56 for 10 at 768p, and $0.49 for 6 seconds at 1080p. A minute made of ten 6-second clips costs $2.80 at 768p and $4.90 at 1080p.

H3 is first on both Artificial Analysis's and Arena's image-to-video leaderboards. It takes 97 to 122 seconds per clip on MiniMax's API, as Artificial Analysis measured on 24 and 25 October. It also has open weights, which we run on one 8 GB RTX 3070 (how we run MiniMax H3 on an 8 GB GPU).

Pika 2.5 pricing

Pika's API costs $0.04 a second at 720p and $0.09 at 1080p, without sound; a soundtrack is a separate pass at $0.005 a second (Pika's API catalogue). Since 5 August 2026 Pika has sold API access through its "API Club": $10 a month, plus what you generate. Text-to-video clips are 5 seconds, so a 10-second 1080p shot from a prompt is two clips, $0.90 before the subscription. Image-to-video clips can run 10 seconds, and the catalogue prices a 1080p image-to-video clip longer than 5 seconds at $0.06 a second. Apart from an image-to-video endpoint on fal, Pika 2.5 is sold only by Pika.

Grok Imagine Video pricing

xAI's Grok Imagine Video 1.5 costs $0.14 a second at 720p and $0.25 at 1080p; Grok Imagine Video 1.5 Lite costs $0.03 and $0.14 (xAI's model pages for 1.5 and 1.5 Lite). Both make clips up to 15 seconds, with optional sound. Grok Imagine 1.5 is in Artificial Analysis's top 10 for image-to-video with audio, and its agent version is fourth on Arena's text-to-video leaderboard.

Other AI video APIs and their prices

Less-searched models, several of them near the top of the leaderboards. Prices are from each maker's own page unless the row says otherwise; where Black Forest Labs' own pages disagree, both prices are shown.

• Two more open-weights models: Tencent's HunyuanVideo 1.5 costs $0.048 a second on Tencent Cloud (30,000 tokens a second at $1.60 per million), in 5-second clips without sound. Kandinsky 6.0 (MIT licence) is sold only on fal: $1.35 (Pro) or $0.16 (Lite) per 5-second clip at its native 480p, and $1.53 or $0.32 upscaled to 1080p, with sound.
• Prices we could not confirm in dollars: Vidu (Q3 Pro, Q3 Turbo and Q4 Preview) works out at $0.06 to $0.12 a second if a credit is worth $0.005, but its live page prices credits in Chinese yuan. PixVerse V6 lists $0.040 a second silent and $0.053 with sound at 720p ($0.080 and $0.102 at 1080p), but a third party implies its credits cost 2.25 times as much. Treat both as unconfirmed.
• No API, or no public price: Utopai X, second on Artificial Analysis's text-to-video leaderboard, is a web app only, at about $1.40 a second in its credits. Meta's Muse Video is consumer-only, and HiDream's O1-Video API has no public price. Midjourney has no public API at all: its terms ban automated access, and the only official sign of one is an "Enterprise API" survey from July 2025.

Does a more expensive AI video model rank higher?

Not on the public leaderboards. Here is Artificial Analysis's text-to-video leaderboard with audio (its September 2026 refresh, read on 25 October), with each model's price per second at 1080p:

The dearest model in the top ten, Seedance 2.5, costs $0.569 a second and ranks fifth, with the same Elo as MiniMax H3, which costs $0.08 a second at 768p. A leaderboard is a vote on its own prompts, which may not be yours, and the boards disagree with each other: Arena's text-to-video board puts Gemini Omni Flash first. (How far to trust a leaderboard is its own question; we traced where 3,778 LLM leaderboard scores come from.)

Cheapest AI video generator, and the cheapest minute of video

It depends on what you need. At list price on 25 October, among the models in the main comparison table:

• 720p, no sound needed: LTX-2.3 Fast, Grok Imagine Video 1.5 Lite and Veo 3.1 Lite without sound ($0.03 a second), Pika 2.5 ($0.04, plus $10 a month), Hailuo 2.3 (about $0.047 at 768p, in 6-second clips) and Luma Ray3.2 ($0.06, in 5-second clips).
• 720p with sound: LTX-2.3 Fast and Grok Imagine Video 1.5 Lite ($0.03, sound optional), Veo 3.1 Lite ($0.05) and MiniMax H3 ($0.08 at 768p).
• 1080p with sound: LTX-2.3 Fast ($0.06), Veo 3.1 Lite ($0.08), Veo 3.1 Fast ($0.12), then Kling 3.0 Turbo, Kling 3.0 Omni and Grok Imagine Video 1.5 Lite ($0.14).
• Long single takes: Wan 3.0 ($0.10 to $0.20 a second), Seedance 2.5 and Kling 4.0 all make 30-second clips, which means fewer joins in a longer piece.
• Smaller makers go lower: Agnes Video 2.5 costs $0.025 a second at 720p, and HeyGen Video 1 $0.015 at 768p until its half-price launch offer ends on 31 October ($0.03 after that).

For a whole minute, multiply by 60 (clip-priced models are filled with their cheapest clip length), and remember that this is one take per shot. A minute of 720p costs $1.80 on LTX-2.3 Fast or Grok Imagine Video 1.5 Lite, or on Veo 3.1 Lite without sound, and $24.00 on Veo 3.1 with sound:

Real projects need retakes. When we budget a test run across many models, we add 10 to 20% for refusals, failures and re-runs, and a shot that takes several tries costs several times its line here.

Run an open-weights model on your own GPU

The cheapest minute of all is one you render yourself, if you have the card and the time. We run MiniMax H3's open weights on one NVIDIA RTX 3070 with 8 GB of memory, through Wan2GP, at 1344x768. It takes about 1.2 to 1.4 GPU-minutes per second of video, so a finished minute is about 72 to 84 minutes of the card's time; the 71.6-second minute we made took 96. There is no fee per second, only electricity. The settings, the speed of each shot and what H3 still gets wrong are in our MiniMax H3 guide.

Check the licence first. MiniMax's licence for H3 does not cover the EU, the UK, South Korea or the USA, which need a separate application. Other open-weights video models and their licences: Wan 2.2 (Apache-2.0, no restriction), Lightricks' LTX-2.x (free under $10 million in revenue), Kandinsky 6.0 (MIT) and HunyuanVideo 1.5 (not valid in the EU, the UK or South Korea). We have run only H3 ourselves.

And if the minute you need is an explainer, with numbers, maps and sources that must be right, a video model is the wrong tool at any price. We make those as code instead (why an explainer needs code).

Resellers: do fal, WaveSpeed, Replicate and OpenRouter charge more?

Mostly not, with some large exceptions. What we found on 25 October, comparing each reseller's listed price with the maker's own:

Three more things matter when you buy through a reseller:

• Coverage. fal and WaveSpeed each sell most video models, but not all: Runway is not on fal, and Marey is only on fal.
• Silent fallbacks. Some services can send a request to a different model. Turn such options off, and check which model answered.
• No batch discount. None of the four offers one on video.

How we collected these prices

• Where: each maker's own price page, read on 25 October 2026 and linked beside every price; a second pass re-read the page of every model in the main table. Where a model is sold only by a reseller (Kling 4.0, Marey and Kandinsky 6.0 on fal), the reseller's page is the source, and the table says so.
• Which price: the list price for API access, billed per second, per clip or per token, plus Runway's and Kling's app plans in their sections. No enterprise deals, batch discounts or promotions, unless a row says so.
• Arithmetic: a script divided clip and token prices down to one second and computed every 10-second, one-minute and plan figure from the stated prices, so the charts and tables use the same numbers. A plan's price per second is its price divided by the seconds its credits buy at the maker's own credit rate, so it assumes every credit is used.
• Not measured here: quality and speed. The rankings and generation times quoted are the leaderboards' and the makers' own.
• Left out: prices that rest on a third party or on a guess, such as Vimoo's (priced in points, with no confirmed dollar value).
• Limits: video prices change often, and some here are promotions (Wan 3.0's 30% off, HeyGen's half-price launch, Kling's first month or year).

We test image models with our own runs: see the best AI image generator in 2026, from 8 models and 96 pictures, GPT Image 2.5 Sunburst against Flare on the same 12 prompts, and how we test AI image generators automatically, with Claude as the judge. If a provider finds a mistake on this page, write to [email protected] and we will correct it.

FAQ

How much does AI video cost per second?

Between about $0.03 and $0.57 a second at list price, at 720p and 1080p, for the best-known models in October 2026. At 1080p with sound, LTX-2.3 Fast costs $0.06 a second, Kling 3.0 $0.168, Veo 3.1 $0.40 and Seedance 2.5 $0.569; the middle of our table is about $0.15. Silent clips and lower resolutions cost less.

What is the cheapest AI video generator?

In our main table, LTX-2.3 Fast, Grok Imagine Video 1.5 Lite and Veo 3.1 Lite without sound cost $0.03 a second at 720p, or $1.80 a minute. With sound at 1080p, LTX-2.3 Fast is the cheapest at $0.06 a second. Agnes Video 2.5 goes lower at 720p ($0.025), and an open-weights model on your own GPU costs only electricity.

Is Sora 2 still available?

No. OpenAI removed Sora 2, Sora 2 Pro, their dated snapshots and the Videos API on 24 September 2026, and named no replacement. Some resellers still list Sora 2, but those listings are stale.

Why is Veo 3.1 so expensive?

Google does not say. Its price page shows that sound doubles the price, from $0.20 to $0.40 a second, and that the full model costs the same at 720p as at 1080p. Veo 3.1 Fast ($0.10 to $0.12 a second with sound) and Veo 3.1 Lite ($0.05 to $0.08) cost far less.

Do AI video generators charge for failed or blocked generations?

Some do. Runway's API documentation says moderated generations cost the same as successful ones, while BytePlus charges nothing when a Seedance video fails, including on content moderation, and Alibaba Cloud does not bill failed Wan requests. Budget 10 to 20% extra for refusals and retries.

Is it cheaper to buy AI video through fal, Replicate or OpenRouter?

Rarely. OpenRouter charged the makers' list prices on everything we checked, plus a 5.5% fee when you buy credits; fal charged about twice BytePlus's price for a US-hosted Seedance 2.5, and Replicate twice Kling's for Kling 3.0. Resellers help where the maker is hard to reach: BytePlus is not available in the United States, and Kling 4.0 is only on resellers so far.

KitCut makes explainer films with Claude. Each film is written as code rather than generated, so the numbers on screen are the numbers in the script.

Make an explainer