Skip to main content

Video generation

Your chatbot can create short video clips on request — a visitor asks for a video in the chat, and the chatbot generates one. Video generation is off by default and fully under your control: you choose the model, what the chatbot is allowed to make, how long each clip can be, and a monthly spend cap.

Video is billed per second of output

A generated video costs far more than a normal chat answer — pricing is per second of output, not per message. Always set a monthly spend cap (below), and keep the maximum clip length short.

Turning it on​

In the dashboard, open your chatbot and find the Video Generation panel, then turn on Enable Video Generation. From there you configure:

SettingWhat it controls
ModelWhich video model powers generation. The dashboard offers a curated shortlist of strong, supported models drawn from the live catalog — for example Veo 3.1 Fast, Veo 3.1, Seedance 2.0, and Wan 3.0 (the exact shortlist can change as the catalog updates). Each option shows its per-second price range (or "token-priced"). If you enable video without picking a model, the chatbot starts with the catalog's default model.
What the chatbot may doThe video operations the chatbot is allowed to use. Today you can turn on Text to video and Image to video; other operations a model advertises — first / last frame, reference to video, video editing, video extension, and motion control — appear marked "Coming soon" and can't be enabled yet. The picker only offers the operations your chosen model actually supports. Veo models are text-to-video only — image-to-video is served by the models that preserve the uploaded frame (Seedance, Wan).
Maximum clip lengthThe longest clip a visitor can request, in seconds. This is the main brake on what a single generation costs, so it's required before you can save an enabled config. Only the clip lengths the model supports are offered, and the visitor picks the actual length for each clip up to this ceiling.
Monthly spend cap (USD)An optional ceiling on total video spend per calendar month (UTC) for this chatbot. Before each clip, its estimated cost is reserved against the cap, so a clip that would take spend over the cap is declined — even if the cap hasn't been reached yet. Leave it empty for no cap.

The panel shows the most a single clip can cost — the model's highest per-second rate at your maximum clip length — so you can see the worst case before turning the feature on. Visitors choose resolution and audio per clip, so most clips cost less. Token-priced models show no estimate.

Model, operations, maximum length, and budget are edited together and saved with Save video settings — because they're interdependent, switching the model re-checks the other choices against what the new model supports. The master on/off switch saves on its own. Resolution, aspect ratio, and audio aren't set here — the visitor chooses them per clip in the chat confirmation (see below).

Video generation is available on every plan — the per-chatbot maximum clip length and monthly spend cap are the cost controls, not your plan tier.

Changes apply to your live chatbot without a redeploy — see How quickly changes take effect.

What visitors experience​

When video generation is on, a visitor can ask the chatbot to make a video in plain language. Depending on the operations you enabled, the chatbot can generate a clip from a text prompt (text to video) or from an image the visitor attaches (image to video) — for example, "animate this photo." An image-to-video clip preserves the uploaded frame. Video comes out in 16:9 or 9:16 only, chosen by the image's orientation — portrait images give 9:16, landscape images give 16:9, and square images come out at 16:9. So a 4:3 photo, being landscape, comes out at 16:9 rather than at its own ratio.

Because each clip costs money, the chatbot asks the visitor to confirm before it generates. The confirmation card shows the full prompt it will use — and, for an image-to-video request, the source frame — alongside the settings for this clip, which the visitor can adjust per video within the limits you set:

  • Length — any supported length up to your maximum clip length.
  • Resolution — the model's supported sizes.
  • Aspect ratio — 16:9 or 9:16, for a text-to-video clip. An image-to-video clip takes its ratio from the image's orientation and can't be changed here.
  • Audio — a native soundtrack, on models that support one.

A live estimated cost updates as the visitor changes these. Every value is re-checked on the server before anything is generated, so a clip is only ever made within your model, length, and budget limits. Nothing is generated, and nothing is billed, until the visitor confirms — and the confirmation stays open for about five minutes. The model may refine that prompt before generating, so the shown wording is the basis for the video rather than an exact, word-for-word script. If a generation doesn't work out, the visitor can try again to re-run the last request.

A few limits apply automatically:

  • A signed-in visitor. Each generation is attributed to a signed-in visitor and the clip is stored privately for them, so generating a video requires a visitor with a signed-in identity. Email verification is not required. On-behalf-of (OBO) callers on the External API can't confirm a video, so video generation isn't available through OBO.
  • One video per message. The chatbot generates at most one clip per turn.
  • The video link lasts 7 days. The finished clip is delivered as a private link that stops working after 7 days, and reopening the conversation later doesn't renew it. Visitors should download any video they want to keep.
  • Length and budget are enforced. Each clip is capped at the maximum length you set, and a clip is declined if its estimated cost would take the month's spend over the chatbot's cap. The cap resets at the start of each calendar month (UTC).

Availability and data residency​

  • EU data residency. Video generation is available on chatbots with EU data residency, but with a residency trade-off to be aware of: the video models run through a shared gateway that pins where a provider runs but not where a request terminates, so an EU chatbot's prompts and generated video may transit a non-EU host. If strict EU-only processing is a requirement for your use case, leave video generation off.
  • Zero data retention. On a zero-data-retention chatbot you can only select a model that is fully covered by ZDR; the dashboard flags a model that isn't.