Skip to content

AI Video Generation API Pricing, Lab by Lab (November 2026)

How AI video APIs bill (per second, per clip, per token, credits), each lab's own rate as published on 1 November 2026, and a batch budget with retries.

Get an API key, pay as you go from $10One key for video, image and voice models. Each call shows its price before it runs.
Get an API key
By Robin MFounder of Aitachyon

Published

Updated

Reading time6 min

Read inEnglishFrançais

AI video generation API pricing looks like one number per model, but the bill is set by four things: the unit the lab charges in, the resolution tier, whether the model adds a soundtrack, and how many takes you throw away before one is usable. This article puts the labs' own rates side by side, read off their pricing pages on 1 November 2026, then turns them into a budget for a real batch with retries counted.

The goal is that you can price a shot list before sending a single request, and decide whether to call a lab directly or through a reseller. I sell these models through Aitachyon, the studio I run, so I say plainly where that route costs more.

Four ways a video API bills you

Per second of output

The simplest unit: a rate times the length you ask for. xAI's models page lists grok-imagine-video-1.5 at $0.080 per second and grok-imagine-video at $0.050. Alibaba's Model Studio pricing (international region) puts wan2.7-t2v and wan2.7-i2v at $0.10 per second at 720P and $0.15 at 1080P. Forecasting is multiplication.

Per clip

MiniMax prices its Hailuo models per video, by resolution and length. On its pricing overview, MiniMax-Hailuo-02 costs $0.28 for a 768P 6-second video, $0.56 for 768P at 10 seconds and $0.49 for 1080P at 6 seconds. The per-second equivalent moves with the step ($0.047 at 6 seconds, $0.056 at 10), so convert before you compare. The newer MiniMax-H3 models on the same page bill per second instead ($0.08 at 768P).

Per token

Google and BytePlus count tokens. Google's Gemini API pricing page (last updated 30 October 2026) bills Gemini Omni 1.1 Flash at $17.50 per million output video tokens, at 5,792 tokens per second of 720p, which it rounds to "approximately $0.10 per second".

BytePlus publishes the formula for Seedance 2.5 on its ModelArk pricing page: tokens = (input video duration + output duration) x width x height x frame rate / 1024, at $10.70 per million tokens for 480p and 720p and $11.70 for 1080p when no video goes in. Its own examples for a 5-second 16:9 clip: $0.514 at 480p, $1.156 at 720p, $2.843 at 1080p. The trap is in the first term: input video counts. Send a 30-second reference and the same 5-second 720p output costs up to $4.838, by the same page.

Credits

Runway sells credits at $0.01 each and charges every model a number of credits per second, per its API pricing page: gen4.5 is 12 credits a second ($0.12), gen4_turbo is 5 ($0.05). Runway also resells other labs' models on the same credits, with minimums: seedance2_5 is 20, 30 or 68 credits a second at 480p, 720p and 1080p, with an 80-credit minimum per generation.

Whatever the unit, convert it to dollars per second of usable footage before comparing anything.

What moves the price

Resolution

Seedance 2.5 on ModelArk costs $0.103 a second at 480p, $0.231 at 720p and $0.569 at 1080p: 5.5 times from bottom to top. Hailuo-02's 6-second clip goes from $0.10 at 512P to $0.28 at 768P and $0.49 at 1080P. For Gemini Omni, Google's Omni guide says 1080p and 4K are upscaled, and the pricing page prices only 720p. My rule: draft at the lowest tier, render the keeper at the resolution the platform will actually show.

Native audio

Some labs bundle sound, some charge for it. Google's Omni guide says the model "will try to generate an appropriate audio track" by default, inside the same token price. Runway's list shows the surcharge where it is separate: veo3.1 is 20 credits a second without audio and 40 with, kling3.0_standard 9 without and 13 with, and Runway's audio parameter defaults to true. If a voiceover and music go on top in the edit, generated sound is money spent twice. What a voiceover costs on the ElevenLabs API covers the narration line.

Model retirements

If you budgeted on Veo 3.1 earlier this year, that line is gone. Google's deprecations page lists the Veo 3.1 preview models shutting down on 22 October 2026, with gemini-omni-1.1-flash as the replacement, and the Gemini API pricing page no longer prices Veo.

Failed renders

Alibaba's page states that failed requests are not billed, and MiniMax says the same for its prepaid video packages. The Google, BytePlus, xAI and Runway pages I read do not say either way. On a batch of hundreds of jobs, ask before you run.

One 6-second clip, lab by lab

The same shot priced on each lab's own API and on Aitachyon, from the published rates on 1 November 2026. Each model name links to its price page.

Model (lab)Lab's unit6 s, lab's own API6 s on Aitachyon
Gemini Omni 1.1 Flash (Google), 720poutput tokensabout $0.61$1.22 to $1.26
Seedance 2.5 (ByteDance, on BytePlus), 720ptokens$1.39$2.77
Wan 2.7 (Alibaba), 720Pper second$0.60$1.20
Hailuo 02 (MiniMax), 768Pper clip$0.28$0.54
Grok Imagine video (xAI)per second$0.30 to $0.48$0.60 to $0.96

Kling is missing on purpose. Kling's developer pricing page builds its table in the browser and I could not read a rate off it on 1 November, so I will not quote one. Runway resells Kling 3.0 Standard at 9 credits a second silent and 13 with audio ($0.54 and $0.78 for 6 seconds). Through my API, Kling v3 is $0.168 a second silent and $0.336 with native audio ($1.01 and $2.02).

Two things stand out. The same model has three prices depending on where you call it: Seedance 2.5 at 720p is $0.231 a second on ModelArk, $0.30 on Runway and $0.4624 through mine. And the model choice moves the price of a shot about five times before resolution or audio even enter.

Get an API key, pay as you go from $10One key for video, image and voice models. Each call shows its price before it runs.
Get an API key

A real batch, with retries counted

Take a launch shot list: 12 product shots of 6 seconds each on Seedance 2.5, delivered at 720p, so 72 seconds of usable footage. The naive budget on ModelArk is 72 x $0.2312 = $16.65. Here is a plan that survives contact with the model:

  1. Drafts. Three takes per shot at 480p, where BytePlus renders its draft mode at the 480p token rate: 12 x 3 x 6 s = 216 seconds x $0.1028 = $22.20.
  2. Finals. One 720p render per shot, plus a second one for half of them: (12 + 6) x 6 s = 108 seconds x $0.2312 = $24.97.
  3. Total. $47.17, or $0.66 per usable second: 2.8 times the 720p sticker rate.

Three drafts and one and a half finals per shot are my planning assumptions, not a statistic. Replace them with your own numbers after the first batch, because the takes-per-keeper ratio moves the total more than any rate in the table. Drafting at 480p is what keeps it sane: a 480p second costs 44% of a 720p one.

The same batch through my API, at $0.2056 (480p) and $0.4624 (720p) a second, comes to $44.41 + $49.94 = $94.35: exactly twice the ModelArk figure, with any failed render refunded to the cent.

How Aitachyon prices the same calls

Aitachyon sells these models through one HTTP API, a hosted MCP server and a web studio, all on one prepaid balance. The rule is public on the MCP page: a call costs what the provider charges, times 2 on video and avatars and times 3 on everything else. The same rule comes back in the prepaid.basis field of GET /api/pricing, a public endpoint that returns every rate as JSON without a key, and each model has its own page under /models.

So the right-hand column of the table is roughly twice the left one, by design. If you run one model at volume, call the lab directly: it is cheaper, and I would do the same. The extra pays for one key across Google, ByteDance, Alibaba, MiniMax, Kuaishou and xAI models instead of six accounts, a failed job refunded to the cent, an optional spend cap per API key, and an agent that can ask what a call will cost before running it. Top-ups start at $10; the details are on the pricing page.

Before you send the batch

  • Pin model, resolution, duration and audio in every request. Defaults cost money: Omni generates a soundtrack by default, and Runway's audio parameter starts at true.
  • Compute the ceiling first: shots x seconds x rate x takes, plus any input video on token-billed models.
  • Check minimum charges, such as Runway's 80-credit floor on seedance2_5.
  • Find out how failed jobs are billed before you run hundreds.
  • Log the cost of every call next to its file, then compute takes per keeper by model.

If an agent will run the batch, generating video from Claude Code over MCP covers the setup, and credits versus pay per call compares the two ways to prepay for it.

Robin MFounder of Aitachyon
How the studio works

One key for every model.

Video, image and voice models behind one API key and one MCP server. Pay as you go from $10: each call shows its price before it runs, and a failed job is refunded to the cent.

Get an API key

How the API and the MCP work