Video generation API pricing, compared per second

Almost every video model bills by the second of video it produces, but the details differ enough to change a bill. The most common form is a price per second that depends on resolution: Alibaba's Wan, xAI's Grok Imagine and MiniMax's H3 family each list separate rates for several output sizes, and the higher tiers cost a multiple of the lowest. Some vendors publish a single flat rate that applies to every resolution, as Runway does for its Gen-4 models.

A second variation is billing by the clip. Luma prices Ray 3.2 per five-second clip, so a shorter or longer request does not scale the way a per-second rate would, and its HDR mode doubles the cost. 1 tracked model is billed this way, and the table converts it to a per-second equivalent. A third is an audio surcharge. Models that generate sound alongside the picture, such as Kling 3.0 through Runway, charge more per second when audio is on, so the silent and with-audio prices are shown in separate columns.

There are also charges that sit outside the headline rate. Image-to-video requests can bill the source image as an input, as MiniMax does, and video-to-video tools such as Runway's Aleph charge a separate, higher rate because they process an input clip as well as producing an output one. Maximum clip length varies from a few seconds to 30 seconds for Wan 3.0, which matters when you plan to stitch shots together.

Every price is converted to dollars per second of generated output. Per-clip prices are divided by the clip length, so a $0.15 five-second clip becomes $0.03 a second. The resolution control maps each model to the nearest published tier: 480p looks for a 480p or 540p price, 720p looks for 720p or 768p, and 1080p looks for 1080p only. A model with a single flat rate is shown at every resolution, labeled as flat so you can tell it is not tier-specific. A model that has neither a matching tier nor a flat rate is left out of the ranking and counted in the line beneath the controls, because estimating a missing tier would be guessing.

Native resolution labels are kept under each price. If a model is priced at 768p and you select 720p, the table says so, so you can see the tier that actually produced the number. The per-clip column multiplies the per-second price by your clip length, and the monthly column multiplies that by clips per day and by 30 days. None of this includes retries, failed generations that are still billed, or subscription fees on the vendor's consumer apps.

Fifty clips of 8 seconds a day is 400 seconds of video a day and 12,000 seconds a month. At 720p, the cheapest model in the table, xAI Grok Imagine Video 1.5 Lite, charges $0.03 a second, so that workload costs $360 a month. The most expensive, ByteDance Seedance 2.5, charges $0.3 a second, which comes to $3,600 a month.

The gap is about 10 times for the same duration of footage. It reflects genuine differences in what the models do, such as native audio, longer clips, stronger motion or video-to-video editing, so the cheapest row is a floor for budgeting rather than a like-for-like substitute for the dearest.

Retries change this arithmetic more than in most categories. Video models are less predictable than text or image models, and it is normal to generate three or four takes to get one usable clip. If your keep rate is one in three, multiply the monthly figure by three. Because every generation is billed whether or not you use it, the most useful number to enter in the calculator is the count of generations, not deliveries.

4 of the 17 rows are sold through a reseller rather than directly by the model's maker. Kling and Seedance are priced here at Runway's rates, and the table marks them with a Reseller tag and a via line. The maker's own price for the same model may differ, and we omit direct Kling pricing because we could not read it reliably. Pika is omitted because its price could not be verified.

Model status moves fast. OpenAI's Sora API was shut down on September 24, 2026, and the Veo 3.1 preview in the Gemini API was shut down on October 22, 2026. LiteLLM's community list still carries entries for both, so we filter them out. We do not show a Veo price: we have not verified one on Vertex AI, and an unverified number is worse than none.

The Confidence column marks every row, and Verified only hides third-party and reseller prices. Some tiers, such as Gemini Omni Flash above 720p, come from Runway's published list rather than Google's page.

Output quality, motion realism, prompt adherence and generation time are not scored. Neither are volume discounts, enterprise plans, free credits or concurrency limits. Compare shortlisted models on your own prompts, since two models at the same price can differ widely on the footage they return.

Prices come from vendor pricing pages where available, tagged Official, and from the LiteLLM community price list, tagged LiteLLM. Reseller and third-party rows are tagged as such. Each row stores its published price basis, source link and the date we last checked it. The dataset was last updated on 2026-10-24 and refreshes daily; values outside plausible bounds for a per-second video price are dropped instead of published.

Not listed: Sora (API shut down on 2026-09-24), Veo (Gemini API preview) (preview shut down on 2026-10-22), Veo (Vertex AI) (general-availability price not verified), Pika (price not verified) and Kling (direct) (vendor price page could not be read; Kling is listed through Runway). For still images, see image generation pricing. If a price here is wrong, use the contact link below and include the model and a source.