Cheapest AI video generation API (2026) — eight models, priced per clip

As of September 2026 the same 8-second 720p clip costs between $0.18 and $3.03 depending on which video model generates it — about a 16.8× spread across the eight production models on one API key. That number, not a quality ranking, is what usually decides which video generation API a team ships on, and it is the one number the roundups on this search do not print, because none of their authors sells more than a couple of these models. This page prints it for all eight, together with the thing that breaks most comparison tables: two of the three providers bill per second of output and one bills per clip, so the cheapest model changes with the length of the clip you generate.

Bias declared: Kunavo is our product — a resold-access gateway, not an independent review site. Every figure below is read live from the catalog that bills the requests, so it is what you would actually be charged rather than a rate copied off a press release. Kunavo is also not the only way to reach these models, and the section on what we do not carry is there because four of the models people ask about are not on this key at all.

What one 8-second 720p clip costs, model by model

Eight seconds at 720p is the common ground, because it is the length a per-clip model bills as one unit and a duration every per-second model supports. Kunavo rates last verified September 14, 2026. The "provider list" column is the vendor's own published price for the same clip: Google's for the Veo 3.1 family, from Google's Veo API documentation and its Gemini API pricing page, and Alibaba Cloud's for Wan 2.7, from its Model Studio video-generation documentation.

Where the provider-list column reads "—", we have no verified published list price from that provider to compare against — ByteDance does not publish a public per-second rate for the Seedance models the way Google and Alibaba Cloud do for theirs — so no saving is claimed for those models. On the 4 models where a list price exists — the Veo 3.1 family and Wan 2.7 — Kunavo's price sits under it. On the 4 ByteDance models it does not follow that the price is better than going direct, and we do not say that it is.

The two billing shapes, and why they break comparison tables

Google's Veo 3.1 models bill one price per clip, whatever its length. The ByteDance Seedance models and Alibaba's Wan 2.7 bill per second of output. Every roundup that prints one number per model has silently picked a duration, and the ranking it produces is only correct at that duration.

The consequence is concrete. At eight seconds, Veo 3.1 Lite at $0.18 leads the table. Halve the clip to four seconds and every per-second model halves with it while the Veo models do not move — so a per-second model that looked expensive at eight seconds can be the cheaper choice at four, and the reverse is true at sixteen. Decide the clip length your product actually generates before you compare prices at all. The AI video cost calculator does this arithmetic across the catalog if you would rather not.

Which model for which job

• Drafts, previews and volume work — Veo 3.1 Lite at $0.18 for eight seconds. At this price a rejected clip costs less than the round trip of asking a human whether to generate it.
• A hero shot that has to hold up — Veo 3.1 Quality. It is the most expensive per-clip option on the key and the one worth paying for when the output ships to a customer.
• Short clips, under about six seconds — a per-second model, because you pay only for what you generate. Seedance 2.0 Mini is the cheapest of them per second.
• Animating art you already have — the Veo 3.1 family or Seedance 2.5, which accept first-and-last-frame and reference-image inputs rather than only a single driving image.
• Text-only prompts at 1080p — Wan 2.7, the one model here with no image-to-video path and a published list price to check ours against.

What Kunavo does not carry

Four of the models most often asked about on this search are not available on Kunavo as of September 2026: Sora 2 and Sora 2 Pro, Kling, Hailuo, and Runway, along with Luma and Pika. If your shortlist has one of those on it, this key will not serve it and a different provider is the correct answer. We keep a page on Sora access and one on Kling that say the same thing and point at what is available instead.

Calling it

One endpoint covers all eight models; the model id is the only thing that changes between them. Generation is asynchronous — the first request returns a task id and the clip arrives on a later poll, which is how every provider in this category works, because a clip takes tens of seconds to render.

# One request, any of the eight models — the model id is the only thing
# that changes. Video generation is asynchronous: this returns a task id.
curl https://api.kunavo.com/v1/videos \
  -H "Authorization: Bearer $KUNAVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "veo-3-lite",
    "prompt": "a paper boat crossing a puddle at dusk, shallow depth of field",
    "resolution": "720p"
  }'

# Then poll the task until it reports a finished clip.
curl https://api.kunavo.com/v1/videos/$TASK_ID \
  -H "Authorization: Bearer $KUNAVO_API_KEY"

Per-model detail lives on the individual guides: Veo 3.1, Seedance, Wan 2.7, and the text-to-video endpoint itself. For the image and music halves of the same key, see the image generation API comparison.

FAQ