Seedance 2.5

Seedance 2.5 is ByteDance's frontier video model for long-form storytelling: a single generation runs up to 30 seconds with synchronized audio, and one request can carry up to 50 reference assets (30 images, 10 videos, 10 audio clips) for character, product, and style consistency. It edits and extends existing clips, follows integer-second timestamps for shot-by-shot direction, and natively understands prompts — and generates speech — in more than 10 languages including Arabic. Output is 480p or 720p at 24fps in MP4 or MOV; for 1080p and 4K delivery, Seedance 2.0 remains available alongside it. Billing is per generated frame, so short drafts stay cheap and you pay exactly for what renders.

Constraints & limits

The gateway auto-clamps requests to the model's limits where possible.

Quick start

Video generation is async — the sample submits then polls until done or failed (auto-refund on failure).

import os, time, requests

API = "https://api.thalam.ai/v1"
headers = {"Authorization": f"Bearer {os.environ['THALAM_KEY']}"}

# 1. Submit
submit = requests.post(
    f"{API}/videos/generations",
    headers=headers,
    json={
        "model": "bytedance/seedance-2.5",
        "prompt": "A drone shot over a desert at golden hour.",
        "duration": 5,
    },
)
task_id = submit.json()["task_id"]

# 2. Poll until done (or auto-refund on failure)
while True:
    poll = requests.get(f"{API}/videos/tasks/{task_id}", headers=headers).json()
    status = poll["task"]["status"]
    if status == "TASK_STATUS_SUCCEED":
        print(poll["videos"][0]["video_url"])
        break
    if status == "TASK_STATUS_FAILED":
        print(f"Failed: {poll['task'].get('reason', 'unknown')}")
        break
    time.sleep(5)