30 seconds, one pass
4–30 seconds generated as a single continuous take, with scene changes and tempo shifts inside it. Not stitched, not extended — there is no seam to hide.
Seedance 2.5 · API live
The Seedance 2.5 AI video generator makes 30 continuous seconds in a single pass — scene changes, tempo shifts and camera moves inside one unbroken take, not stitched and not extended. Generate it in the browser, or with one API call.
30 seconds in a single pass — no stitching, no seam to hide.
Seedance 2.5 is ByteDance’s video-generation model, live on the BytePlus ModelArk API since 16 July 2026. It generates up to 30 seconds of continuous video in a single pass — scene changes and tempo shifts inside one take — and accepts up to 50 multimodal references. This site resells it with the per-second cost shown before you generate.
Earlier Seedance models — 2.0, 2.0 Fast, 2.0 Mini and 1.5 Pro — top out at 15 seconds and reach longer runs by stitching or extending, which is where continuity usually breaks. 2.5 removes the join entirely. That is the whole reason to pay more per second for it.
Figures ByteDance has announced but we have not yet confirmed against the API are marked unverified on the model page rather than repeated as fact here.
4–30 seconds generated as a single continuous take, with scene changes and tempo shifts inside it. Not stitched, not extended — there is no seam to hide.
50 multimodal inputs in one generation: 40 images, 5 video clips and 5 audio tracks, against roughly a dozen on 2.0.
Change one element and leave the rest of the frame untouched, instead of regenerating the whole take and hoping the rest survives.
Green-screen and white-model references drive motion, spatial layout and character interaction, so generation follows a structured visual plan rather than text alone.
Plan the camera move against a white-box preview before spending a generation on it. The cheapest take is the one you did not need to run.
Dialogue in multiple languages with lip movement synced to it, carried by the model rather than added in post.
Every figure marked unverified is ByteDance’s own announced claim, which we have not yet confirmed against the provider API. No independent benchmarks for Seedance 2.5 existed when this page was written — so rather than restate a keynote, we mark the gap and publish measurements as they land.
Every take on this page was generated with Seedance here — no stock, no cherry-picked vendor reel. Prompt, model and settings shown on each one.
Nothing published yet. This grid fills with takes generated here, each reviewed before it goes public.
Every Seedance 2.5 capability figure in circulation is ByteDance’s own. Our benchmark measures what actually comes back — billed cost, latency, failure rate, and how often a 30-second take holds character end to end. It has not been run yet. The rates below are the published vendor figures our pricing currently rests on, not measurements, and the columns we cannot honestly fill are marked as such.
Billing shapes differ — some routes charge for input-video seconds, some add a per-request token component — so headline per-second rates are not comparable until they are normalised. These figures are.
For developers
Idempotency keys on create, rate-limit headers on every response, a stable error code you can switch on, and a status page publishing real per-model p50 and p95. Prepaid packs never expire.
curl https://sedance.app/api/v1/generations \
-H "Authorization: Bearer sk_live_..." \
-H "Idempotency-Key: $(uuidgen)" \
-H "Content-Type: application/json" \
-d '{
"model": "seedance-2.5",
"duration_sec": 30,
"resolution": "1080p",
"prompt": "one continuous take, camera rising over a night market",
"confirm_cost": true
}'
→ 202 { id, status: "reserved",
credits_held: 7179, eta_seconds: 210 }