Seedance 2.5 (2026): Release Date, Specs and API Code

Seedance 2.5 is ByteDance’s video generation model that turns a text prompt or a still image into a clip of 4 to 30 seconds, with sound generated in the same pass. This guide covers what it is and when it came out, its exact specs, what a clip costs, working API code, how it compares with the other video models you are probably weighing, and the mistakes that burn credits for no reason.

What is Seedance 2.5?

Seedance 2.5 is the latest model in ByteDance’s Seedance line, which started in 2025 and drew wide attention with Seedance 2.0’s short cinematic clips. It is a joint audio-video model: instead of producing silent footage and leaving you to add sound later, it generates the picture and a matching soundtrack together, so a door slam lands when the door actually closes.

It handles the two jobs most people need from a video model:

• Text-to-video: you describe a scene and the model creates it from nothing.
• Image-to-video: you supply a start frame (a product photo, a character design, a still from your own footage) and the model animates it. You can also supply an end frame so the clip finishes on a specific image.

Seedance 2.5 release date

Seedance 2.5 was released on 25 June 2026. You will find other dates floating around, because previews, announcements and wider rollouts of video models rarely happen on a single day. For practical purposes, it has been generally available to developers since late June 2026, including on Unicorn API under the model id bytedance/seedance-2-5.

What changed from Seedance 2.0

The headline change is length. Seedance 2.5 produces a single clip of up to 30 seconds in one generation, without stitching shorter clips together, which matters because stitched clips tend to drift: a character’s jacket changes colour, the lighting jumps, the camera resets. It also accepts a larger set of reference media (images, video and audio) to steer the result, plus an optional web search step that grounds the prompt in current information. If you are still running the older model, the Seedance 2.0 pricing guide covers how that one is billed.

Seedance 2.5 specs: length, resolution, audio and inputs

Here is what the model supports when you call it through Unicorn API.

Spec	Seedance 2.5
Modes	Text-to-video, image-to-video (start frame, optional end frame)
Duration	4 to 30 seconds, in 1-second steps
Resolution	480p, 720p, 1080p
Aspect ratios	16:9, 9:16, 1:1, 4:3, 3:4, 21:9, adaptive
Audio	Native, generated with the video (with_audio)
References	Reference images, videos and audio by URL
Extras	Optional web search, optional return of the last frame
Price on Unicorn API	315 credits per video, flat

The input fields, explained

• prompt: the scene description. The only field you strictly need.
• image and end_image: URLs for the first and last frames. Use both when you need the clip to start and end on known shots, such as a product reveal.
• duration, resolution, aspect_ratio: length in seconds, output quality and frame shape. Adaptive means the output follows the shape of your start image, so a vertical photo produces a vertical video.
• with_audio: turns native sound on or off.
• reference_image_urls, reference_video_urls, reference_audio_urls: media that guides style, motion, characters or sound without being used as a literal frame.
• web_search: lets the model look things up to ground the prompt, useful when your prompt mentions something recent that may not be in its training data.
• return_last_frame: asks for the clip’s final frame, which is handy when you want to continue a scene in a second generation.
• mode and output_format: advanced controls; check the API documentation for accepted values before you set them.

One thing the specs do not include is 4K. Some sites advertise the Seedance 2.5 AI video generator as a 4K tool, but on Unicorn API the ceiling is 1080p. If you need 4K, that is a reason to look at Kling 3.0 Omni, covered in the comparison below.

How much does Seedance 2.5 cost?

On Unicorn API, every Seedance 2.5 video costs a flat 315 credits. Duration and resolution do not change the price: a 4-second 480p test and a 30-second 1080p clip with audio cost exactly the same.

In dollars, that depends on how you buy credits:

• Monthly plans (Starter $25, Creator $50, Studio $100, Pro $200) give credits at $0.01 each, so one clip costs $3.15.
• Top-ups cost $0.015 per credit and never expire, so one clip costs about $4.73.

The full breakdown is on the pricing page. Because the price is flat, the effective cost per second falls sharply as clips get longer:

Duration	Credits per clip	Credits per second	Cost per clip on a plan
4 s	315	about 78.8	$3.15
10 s	315	31.5	$3.15
15 s	315	21	$3.15
30 s	315	10.5	$3.15

The practical takeaway: if your final asset is long, Seedance 2.5 is very good value, because a 30-second clip costs about 7.5x less per second than a 4-second one. If you only ever need short clips, a per-second model will usually be cheaper. On a Starter plan’s 2,500 credits, you get 7 full Seedance clips with credits left over.

You never have to guess. Send the same request body to POST /v1/models/price and you get the exact charge before anything runs, which is worth building into any app where users choose the settings.

How to use the Seedance 2.5 API (with code)

Video generation takes longer than a normal HTTP request should stay open, so Unicorn API runs it as a job: you create the job, then poll it until it finishes and read the output URL. The flow is the same for every video model on the platform, so once this works, switching models is a one-line change.

Step 1: get a key

Create an API key on the keys page and export it as an environment variable:

export UNICORN_API_KEY="your-key-here"

Step 2: check the price

curl -X POST https://api.unicornapi.net/v1/models/price \
  -H "Authorization: Bearer $UNICORN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"bytedance/seedance-2-5","input":{"prompt":"A slow drone shot over a rocky coastline at sunset, waves breaking","duration":30,"resolution":"1080p"}}'

Step 3: generate a text-to-video clip in Python

This script prices the request, creates the job, polls until it is done and prints the video URL.

import os
import time
import requests

API = "https://api.unicornapi.net"
HEADERS = {"Authorization": f"Bearer {os.environ['UNICORN_API_KEY']}"}

body = {
    "model": "bytedance/seedance-2-5",
    "input": {
        "prompt": (
            "A slow drone shot over a rocky coastline at sunset, waves breaking "
            "against dark rocks, gulls calling, wind and surf audible"
        ),
        "duration": 30,
        "resolution": "1080p",
        "aspect_ratio": "16:9",
        "with_audio": True,
    },
}

# 1. Know the cost before you spend it
price = requests.post(f"{API}/v1/models/price", json=body, headers=HEADERS)
price.raise_for_status()
print("Price:", price.json())

# 2. Create the job
resp = requests.post(f"{API}/v1/jobs", json=body, headers=HEADERS)
resp.raise_for_status()
job = resp.json()
print("Job created:", job["id"])

# 3. Poll until it finishes (give up after 15 minutes)
deadline = time.time() + 15 * 60
while job["status"] not in ("succeeded", "failed"):
    if time.time() > deadline:
        raise TimeoutError(f"Job {job['id']} still {job['status']}")
    time.sleep(5)
    r = requests.get(f"{API}/v1/jobs/{job['id']}", headers=HEADERS)
    r.raise_for_status()
    job = r.json()

if job["status"] != "succeeded":
    raise RuntimeError(f"Job failed: {job}")

# 4. Read the result
print("Video URL:", job["outputs"][0]["url"])

Step 4: animate an image instead

For image-to-video, add a start frame and, optionally, an end frame. With "aspect_ratio": "adaptive", the output matches the shape of your start image.

curl -X POST https://api.unicornapi.net/v1/jobs \
  -H "Authorization: Bearer $UNICORN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bytedance/seedance-2-5",
    "input": {
      "prompt": "The camera slowly circles the sneaker on a concrete plinth, soft studio light, a quiet electronic beat",
      "image": "https://example.com/sneaker-front.jpg",
      "end_image": "https://example.com/sneaker-side.jpg",
      "aspect_ratio": "adaptive",
      "duration": 8,
      "resolution": "1080p",
      "with_audio": true
    }
  }'

Then poll GET /v1/jobs/{id} exactly as in the Python script. If you would rather try prompts before writing code, the video studio runs the same models in the browser, and the agents page shows how to call them from Cursor or Claude Code.

Seedance 2.5 vs Wan 3.0, Kling 3.0 Omni and Grok Imagine

The Seedance 2.5 AI video generator is not the cheapest option or the highest-resolution one. Its strength is long single shots with native audio at a predictable price. Here is how it lines up against the other video models on Unicorn API, using credits per clip at 1080p.

Model	Modes	Max length	Max resolution	Audio	10 s at 1080p	30 s at 1080p
Seedance 2.5	Text, image	30 s	1080p	Native	315	315
Wan 3.0	Text, image	30 s	1080p	Native, no extra cost	320	960
Kling 3.0 Omni	Text, image, reference, transform	15 s	4K	Native	70	Not available
Grok Imagine Video 1.0	Text, image	30 s	1080p	Not listed	80	240
Grok Imagine Video 1.5	Image only	15 s	1080p	Not listed	80	Not available

When Seedance 2.5 is the right pick

• You need one continuous shot longer than 15 seconds, with sound, and consistency matters more than squeezing the price.
• You want to steer with references: style images, a motion reference video, or an audio reference.
• You want a cost that does not move when someone changes the duration slider.

When something else fits better

• Wan 3.0 is cheaper for short or lower-resolution clips. At 1080p it bills 32 credits per second, so it undercuts Seedance below about 10 seconds. At 720p it stays cheaper up to 19 seconds, and at 480p even a 30-second clip is only 240 credits. Past those points, Seedance wins.
• Kling 3.0 Omni is the choice for 4K, storyboarded multi-shot sequences and character elements, at a flat 70 credits, as long as 15 seconds is enough.
• Grok Imagine Video 1.0 and 1.5 are the budget options for silent or draft footage. A 30-second 480p clip on Grok 1.0 costs 72 credits; the Grok Imagine 15-second guide covers both in detail.

Since everything runs behind one key and one bill, you do not have to commit to a single model. The full model catalog lists the rest, and changing "model" in the request body is all it takes to switch.

How to write better Seedance 2.5 prompts

A long clip gives the model more time to wander, so prompt structure matters more at 30 seconds than at 5. These habits help:

1. Lead with the camera. “Slow dolly in”, “static wide shot”, “handheld follow” tell the model how the frame should move, which is the single biggest factor in whether a clip feels deliberate.
2. Describe the subject once, precisely. “A woman in a yellow raincoat with a black umbrella” beats “a woman” followed by details scattered across the prompt. Precise descriptions reduce drift over long durations.
3. Write the action as a sequence. For a 20 or 30-second clip, describe what happens in order: “she walks to the edge of the pier, stops, then turns to face the camera as the rain eases.”
4. Say what you want to hear. Native audio follows the prompt. Name the sounds: “rain on the umbrella, distant traffic, a foghorn.” If you say nothing, you get whatever the model infers.
5. Use references instead of adjectives. A single reference image of the look you want often does more than “cinematic, moody, high contrast”.
6. Turn on web search only when it helps. It is for prompts that depend on current facts, such as a recently launched product’s appearance. For a generic scene, it adds nothing.

A worked example

Say you need a 30-second vertical ad for a coffee brand. A weak prompt is “coffee shop, cosy, cinematic”. A stronger one:

Vertical handheld shot. Morning light through a café window. A barista in a green apron pours steamed milk into a white cup, forming a leaf pattern. The camera pushes in slowly to the cup, then tilts up as a customer’s hand lifts it. Sounds: milk steaming, cups clinking, soft chatter.

Send it with "duration": 30, "aspect_ratio": "9:16", "resolution": "1080p" and "with_audio": true. The cost is 315 credits, or $3.15 on a plan. The same clip on Wan 3.0 at 1080p would be 960 credits, so for this brief Seedance 2.5 is about 3x cheaper.

Common Seedance 2.5 mistakes (and how to avoid them)

• Testing at short, low-res settings to “save credits”. With flat pricing, a 4-second 480p test costs the full 315 credits. Either test at your real settings, or draft the idea on a per-second model like Grok Imagine Video 1.0 at 480p, then render the final on Seedance. Remember a different model will interpret the prompt differently, so treat drafts as rough checks of composition, not previews.
• Fighting the start frame. In image-to-video, forcing 16:9 on a portrait photo means the model has to invent or crop content. Use adaptive unless you have a reason not to.
• Assuming audio is on. Set with_audio explicitly so you are never surprised by a silent clip, or by sound you did not want.
• Expecting 4K. 1080p is the maximum here. Plan to upscale, or use Kling 3.0 Omni for short 4K shots.
• Cramming several scenes into one prompt. One continuous shot with a clear progression works far better than “first a beach, then a city, then a forest”. For multiple scenes, run multiple jobs and use return_last_frame to carry continuity from one into the next.
• Polling too aggressively or forever. Poll every few seconds, set a timeout, and handle a failed status so one stuck job does not hang your pipeline.
• Trusting third-party spec claims. You will read about 3-minute clips and huge reference sets. Build against what the API actually accepts, which the docs and the model page list.

Should you use Seedance 2.5?

If your work involves shots longer than 15 seconds that need sound, Seedance 2.5 is one of the most practical options available: 4 to 30 seconds per clip, up to 1080p, native audio, start and end frame control, and one fixed price of 315 credits whether you ask for 4 seconds or 30. If you mainly make short clips, pair it with a per-second model and route each request to whichever is cheaper for that length.

The fastest way to judge it is to run your own prompt. Create a key, price a 30-second request, and compare the output against Wan 3.0 or Kling 3.0 Omni using the same body with a different model id. For more on the Seedance family, the Seedance 1.0 Pro Fast guide covers the faster, lighter end of the line.