Short answer, updated 8 August 2026: the Seedance 2.5 API is live on apimodels.app. Pass model "seedance-2.5" to the video endpoint — 480p $0.134/s, 720p $0.300/s, charged only on success and refunded on failure. No Volcengine account and no enterprise verification: sign up, create one API key, call it. This page said the opposite until today, and that was accurate when we wrote it — probing our production Volcengine Ark account on 17 July and again on 4 August, every Seedance 2.5 model id returned InvalidEndpointOrModel.NotFound while Seedance 2.0 returned a task id on the same request. That has changed: doubao-seedance-2-5-260628 now resolves on the same account, so we shipped it. ByteDance's official prompting guide for 2.5 is walked through in English at /blog/seedance-2-5-prompting-guide, and the sample below is a real, working call.
What Seedance 2.5 actually does: ByteDance unveiled it on 23 June 2026 at its Volcano Engine FORCE conference, skipping several intermediate versions. What the shipped API exposes — per ByteDance's model documentation and our own calls — is 480p and 720p output (there is no 1080p tier and no 4K tier), 4 to 30 seconds or duration -1 to let the model choose the length, aspect ratios 21:9 / 16:9 / 4:3 / 1:1 / 3:4 / 9:16 and adaptive (the default), up to 50 reference assets in one job (30 images, 10 video clips and 10 audio clips, each clip 2-30s, with all videos together and all audios together capped at 30s), mp4 or mov output, native synced audio, a web search tool, real people via the asset:// library, and eleven languages natively — Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese and Korean. Two new task types arrive with it: video editing (replace, add or remove objects, or edit the audio track) and video extension (forwards or backwards, over multiple rounds). The FORCE announcement also claimed native 4K at 10-bit; that is not exposed on the API we call, so treat it as an announcement figure rather than something you can request.
Three constraints worth knowing before you write code, because upstream enforces them asynchronously and we do not: video editing, video extension and first-frame / first+last-frame generation all force aspect_ratio "adaptive" — the output keeps the source's ratio and you cannot override it. Video editing additionally forces duration -1, so the output keeps the source clip's length (it may come back up to 0.4s shorter). The source video for an edit must itself be 4-30 seconds. Break any of these upstream and you get InvalidParameter.TaskTypeConstraint after the task has already queued; our API rejects the same request at create time so you get an immediate 400 instead.
What it costs: 480p $0.134/s and 720p $0.300/s of output, charged only on success. Underneath, Volcengine settles on real token usage — tokens = (input video seconds + output seconds) × width × height × fps ÷ 1024 — and that carries one caveat we would rather you heard from us. A job that carries a reference video is billed on its input seconds as well as its output seconds; the per-token rate is lower, but break-even sits at input = output × 0.667, so editing or extending a clip as long as your output costs about 20% more than generating the same length from scratch. It is not a discount tier. And if you need 1080p, Seedance 2.5 does not have it — Seedance 2.0 stays listed on the official channel at 480p $0.092/s, 720p $0.197/s, 1080p $0.492/s, 4-15 seconds.
cURL
# Seedance 2.5 — live since 2026-08-08. 480p $0.134/s, 720p $0.300/s.
# 480p/720p only (no 1080p, no 4K); duration 4-30, or -1 to let the model choose.
curl https://api.apimodels.app/v1/video/generations \
-H "Authorization: Bearer $APIMODELS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "seedance-2.5",
"prompt": "a hummingbird hovering near a bright flower, slow motion",
"resolution": "720p",
"duration": 5
}'
# -> { "data": { "taskId": "..." } }, then poll:
# curl "https://api.apimodels.app/v1/video/generations?task_id=..." -H "Authorization: Bearer $APIMODELS_API_KEY"Python
import requests, time
H = {"Authorization": "Bearer YOUR_KEY", "Content-Type": "application/json"}
B = "https://api.apimodels.app/v1/video/generations"
# Seedance 2.5 — 480p/720p only, duration 4-30 (or -1 = the model picks the length).
# Video editing / extension / first-frame jobs force aspect_ratio "adaptive", and
# editing also forces duration -1; overriding either is a 400 here, not a queued failure.
task = requests.post(B, headers=H, json={
"model": "seedance-2.5",
"prompt": "a hummingbird hovering near a bright flower, slow motion",
"resolution": "720p", "duration": 5,
}).json()["data"]["taskId"]
while True:
r = requests.get(B, headers=H, params={"task_id": task}).json()["data"]
if r["state"] in ("completed", "failed"):
print(r); break
time.sleep(5)Yes, since 8 August 2026. Pass model "seedance-2.5" to https://api.apimodels.app/v1/video/generations — 480p $0.134/s, 720p $0.300/s, charged only on success. This is a change from what this page said before: on 17 July and again on 4 August, every Seedance 2.5 model id we probed on our production Volcengine Ark account returned "InvalidEndpointOrModel.NotFound", which is why we refused to list it. The id doubao-seedance-2-5-260628 now resolves on that same account, so it is live.
It is out. ByteDance unveiled it on 23 June 2026 at its Volcano Engine FORCE conference; BytePlus coverage cited an API date of 16 July that never materialised for third-party accounts like ours, and our 4 August probe still came back empty. It became callable on our Ark account shortly after, and we listed it on 8 August 2026 as seedance-2.5.
On the shipped API: 4-30 second clips against Seedance 2.0's 4-15, plus duration -1 to let the model pick the length; up to 50 reference assets in one job (30 images, 10 videos, 10 audios) against Seedance 2.0's 9 images plus a reference video and audio, and 2.5 accepts an audio-only reference where 2.0 needs an image or video alongside it; two new task types, video editing and video extension; and mov output alongside mp4. Resolution goes the other way — Seedance 2.0 standard reaches 1080p, Seedance 2.5 tops out at 720p. The FORCE announcement claimed native 4K at 10-bit, but the API exposes 480p and 720p only. Native synced audio, the web search tool and asset:// real-person references are on both.
Sign up at apimodels.app, create one API key, and pass model "seedance-2.5" — no Volcengine account and no enterprise verification. The same key already calls Seedance 2.0, VEO, Kling, Sora and the rest, on the same endpoint with the same create-task-then-poll flow. If you already built against seedance-2.0, moving over is a one-string change with two things to check: 2.5 has no 1080p, and its duration range is 4-30 rather than 4-15.
Per second of output: 480p $0.134/s and 720p $0.300/s, charged only on success and refunded on failure. There is no 1080p or 4K tier to price. Settlement is on real token usage, and Volcengine's formula is tokens = (input video seconds + output seconds) × width × height × fps ÷ 1024 — so a job carrying a reference video is billed on its input seconds too. The per-token rate is lower in that case, but break-even sits at input = output × 0.667, which means editing or extending a clip as long as your output costs about 20% more than generating the same length from scratch. It is not a discount tier. For comparison, our official-channel Seedance 2.0 is 480p $0.092/s, 720p $0.197/s, 1080p $0.492/s — Seedance 2.5 at 720p is roughly 52% more.