Video generation
Submit a job, poll it, fetch the result from our own domain. Billed on the output tokens a video is worth, which scale with resolution, frame rate and duration.
A video is a job, not a reply
Every other endpoint here answers your request. This one accepts it: the call returns in under a second with an id, the work takes one to two minutes, and you poll for the result. Nothing streams, and there is no request you can hold open until the video exists.
The id you get back is ours. It is the only handle you need, and the only one we will ever quote you.
Submit
MODEL_ID rather than naming a model, because which ids you can send depends on your key: GET /v1/models with your key is the list that can actually be called, and your models page shows the same thing with prices. The table further down lists what each published model takes — some of those are announced ahead of serving, so the id list is the one to trust before a first call.curl https://api.tokenify.dev/contents/generations/tasks \
-H "Authorization: Bearer $TOKENIFY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "MODEL_ID",
"resolution": "480p",
"duration": 4,
"content": [{"type": "text", "text": "A paper boat on still water"}]
}'content is a list, the way a chat message's content is. A text item is your prompt. An image_url item is a still to animate, and a video_url item is a clip to extend — each only on the models that accept them, which the table below states per model.
Poll
curl https://api.tokenify.dev/contents/generations/tasks/TASK_ID \
-H "Authorization: Bearer $TOKENIFY_API_KEY"status is queued, running, succeeded, failed or cancelled. On success, content.video_url is a link on our own domain — we fetch the finished video from the supplier and serve it ourselves, so the URL you get never points anywhere else.
Five seconds between polls is plenty. A job that has not finished in thirty minutes is given up on and refunded.
What it costs
A video is billed on output tokens only. There is no prompt charge: the count scales with the pixels and the seconds, roughly width × height × fps × duration ÷ 1024, so doubling the resolution costs about four times as much and doubling the length about twice.
You are charged what the supplier reports, bounded by what your request reserved. A job that reports no usage at all is charged the estimate it was accepted on, and a job that fails is not charged.
Two things worth knowing before you build on it
A job cannot be cancelled once accepted. Deleting it removes the record and the stored video; it does not stop the work, and it does not refund it. Submitting is the commitment.
Concurrency is the limit that binds, not requests a minute. A handful of jobs run at once per model across everyone we serve, and a job takes one to two minutes — so the useful figure is how many you can have in flight, not how many you can send.
Past that limit your job waits rather than being refused. Jobs queue in one line shared by every caller, in the order they arrived, and each is sent to the model as a slot frees. You get the same 200 and the same job id either way; poll it and the status reads queued until it starts. There is nothing to retry and nothing to back off from.
A job that waits too long is given up on and refunded in full, with queue_wait_exceeded in its error field. Your own account and key limits are still refusals, with 429 and video_jobs_in_flight: those bound how many jobs you may have outstanding at once, queued ones included.
The stored video is deleted after seven days. Fetch it and keep your own copy if you need it longer. Deleting a job removes the video from our storage immediately, but the link is served with a long cache lifetime — so a browser or proxy that already fetched it may keep showing its own copy for a while afterwards. Our storage is the authoritative one; a cache we do not control is not.
The models, and what each one takes
| Model | Resolutions | Length | Inputs |
|---|---|---|---|
bytedance/seedance-2.0 | 480p, 720p, 1080p, 4k | 4–15s | image, text |
bytedance/seedance-2.0-fast | 480p, 720p | 4–15s | image, text |
bytedance/seedance-2.0-mini | 480p, 720p | 4–15s | image, text, video |
bytedance/seedance-2.5 | 480p, 720p, 1080p | 4–30s | image, text, video |
bytedance/seedance-2.5-flash | 480p, 720p | 4–30s | image, text, video |
These lists are exhaustive. A resolution or a duration outside them is refused before anything is charged, and a reference input on a model that does not take one is refused the same way.
Errors
The shapes are the same as everywhere else here — see Errors. The ones specific to this endpoint are invalid_resolution, video_jobs_in_flight, running_task_deletion for a delete of a job that has not finished, and three that arrive on a polled job rather than as a response status: queue_wait_exceeded, submit_outcome_unknown and dispatch_attempts_exhausted. Each of those three means nothing was generated and the reservation was released in full.
Last updated 2026-11-16.