AI video generation API price comparison (2026) — 8 models, based on one 8-second 720p clip

The same 8-second 720p video can cost anywhere from $0.18 to $3.03 — approximately 16.8 times as much — depending on the model used. This is based on 8 video models available under one API key. But there is one thing you need to know before that number: these 8 models do not use the same billing method — 3 charge a fixed price per video, while 5 charge per second of output. Therefore, the “cheapest model” changes according to the length of the clips you actually make. A comparison table with one number per model shows a ranking that is accurate only for the assumed length without saying what that length is.

First, full disclosure. Kunavo is our product and a gateway that resells model access, not an independent review site. The figures below are read when the page renders from the catalog that actually bills requests, so they are not rates transcribed from press releases but amounts actually charged. Final Kunavo rate check: October 1, 2026.

Price of one 8-second 720p video, by model

We use 8 seconds at 720p as the common benchmark because it is the billed unit length for per-video models and a supported length for every per-second model. The “Provider list price” column shows the price publicly stated by the provider under the same conditions — for Veo 3.1 models, in Google's Veo API documentation; for Wan 2.7, in Alibaba Cloud Model Studio documentation.

A row whose provider list price is “—” means we could not verify a published list price from that provider. Therefore, we make no claim whatsoever about savings for those models. For the 4 models with a published list price (Google Veo 3.1 Lite·Fast·Quality, Alibaba Wan 2.7), Kunavo's price is below it. For the 4 ByteDance models without a list price, this does not mean they are cheaper than using them directly, and we do not say that. We chose to leave the cells blank rather than fill them with estimates.

The two billing methods break comparison tables

The Google Veo 3.1 family charges a fixed fee per video, and a video is always 8 seconds long. It is the same whatever length the request asks for. The ByteDance Seedance family and Alibaba Wan 2.7 charge per second of output. Any comparison table that lists one number per model is quietly assuming a particular length, and its ranking is only correct for that length.

The result is specific. At 8 seconds, Veo 3.1 Lite is at the top of the table at $0.18. If you ask for 4 seconds, all per-second models become half price, while the Veo family still makes an 8-second video and charges the price of one video — so a per-second model that looked expensive at 8 seconds may be the cheaper choice at 4 seconds. Past 8 seconds, the Veo family needs one more whole video while per-second models charge only for the extra seconds, so per-second billing gains ground again. At 16 seconds every model costs twice its 8-second price, and the ranking is the same as at 8 seconds. Decide the clip length your product actually produces before comparing prices. The AI video cost calculator does that arithmetic for you.

Which model to use for each purpose

• Drafts, previews, and high-volume work — Veo 3.1 Lite, $0.18 for 8 seconds. At this price, one disposable shot costs less than a round trip asking someone, “Should we generate this?”
• A single shot going to a customer — Veo 3.1 Quality ($1.60 for 8 seconds). It is the most expensive per-video option, and when the output is going straight to the customer, it is worth paying for.
• Short clips of around 6 seconds — a per-second model. You pay only for what you make. The lowest per-second price among them is Seedance 2.0 Mini ($0.0526 / second).
• When animating existing art — Google Veo 3.1 Lite·Fast·Quality, ByteDance Seedance 2.0 Mini·2.5. It accepts not only a driving image, but also first- and last-frame specifications and reference images.
• 1080p from text alone — Wan 2.7. It is the only option here without an image-to-video path, and it is also the model with a published list price available for comparison.

What Kunavo does not offer

First, here are the models that other pages in these search results list but Kunavo does not offer. As of September 2026, Sora, Kling, Hailuo, Runway, Luma, Pika is unavailable. If one of these appears in the candidate list, it cannot be called with this key, and another provider is the correct answer. Kunavo has separate pages for Sora access and Kling, which make the same point and explain what is available instead.

How to make a call

8 models use one endpoint. The only thing that changes between models is the model ID. Generation is asynchronous — the first request returns a job ID, and the completed video is retrieved through subsequent polling. Rendering one video takes tens of seconds, so every provider in this space uses the same structure. Request formats and polling are covered in the video endpoint documentation, while model-specific details are in the Veo 3.1, Seedance, and Wan 2.7 guides. For image generation with the same key, see the image generation API comparison.

Frequently asked questions