Seedance 2.5 in Practice: Capabilities, API Workflow, and Cost

The short version

Seedance 2.5 is ByteDance Seed’s audio-video generation model for longer, reference-driven clips. It supports 4–30 second videos, up to 50 model-level image, video, and audio references, multi-round extension, and several editing modes.

As of September 3, 2026, the model is available for text-to-video and image-to-video through both CometAPI and BytePlus. The services are separate: they use different credentials, endpoints, model identifiers, and billing systems.

The API workflow is asynchronous. Submit a task with POST /v1/videos, store the returned task ID, and poll GET /v1/videos/{id} until the render finishes.

What Seedance 2.5 is designed to do

Seedance 2.5 is a joint audio-video model rather than a simple text-to-video system. It can:

• Generate a complete clip from a text prompt
• Animate a reference image
• Use image, video, and audio references
• Guide subject identity, movement, camera language, atmosphere, and sound
• Extend a sequence across multiple rounds
• Apply more targeted edits to an existing creative direction

That makes it more useful for reference-led production than for isolated visual experiments. Possible applications include campaign concepts, product stories, social clips, education, and previsualization.

I would not treat the model’s feature list as a guarantee of production quality. Prompt adherence, subject consistency, audio synchronization, physical plausibility, queue time, and cost per usable clip still need to be measured with your own material.

The changes that matter

Longer native generations

Seedance 2.5 generates clips from 4 to 30 seconds. A 30-second ceiling is useful when a sequence needs several story beats, camera movements, or transitions. Instead of stitching many short generations together, a team can attempt more of the sequence in one render.

That does not eliminate continuity problems, but it can reduce the amount of repair work between shots.

More reference inputs

ByteDance documents support for up to:

• 30 images
• 10 videos
• 10 audio clips

These references could represent product angles, character sheets, approved brand imagery, motion examples, camera references, music, or voice cues.

The model-level limit is not necessarily the same as the limit exposed by every API. Check the endpoint documentation before building a workflow around a particular combination of reference types.

Extension and targeted editing

The model adds multi-round extension along with reference-based, timestamp-level, green-screen, and camera-perspective editing controls.

ByteDance also highlights:

• Clay-render references
• Motion references
• Stronger scene continuity
• Multi-round extension

These features are promising for revising a shot without restarting from an entirely new concept. However, platform integrations may expose only some of them, and they should not be confused with a universal quality score.

Availability and model identifiers

Seedance 2.5 was officially released on July 31, 2026. As of September 3, 2026, it is listed for text-to-video and image-to-video, with documented output from 4 to 30 seconds at 480p or 720p on the route described below.

BytePlus documents dreamina-seedance-2-5-260628 for its asynchronous enhanced video API.

There is an identifier discrepancy in the other provider’s documentation:

• The live model page and request example use seedance-2-5
• One API support table shows seedance-2-5-260628
• BytePlus uses dreamina-seedance-2-5-260628

Before hard-coding a production configuration, query:

GET /v1/models

Use the model ID returned for your account rather than assuming that every documentation page is synchronized.

Calling the asynchronous API

The request endpoint is:

https://api.cometapi.com/v1/videos

The request uses multipart form data. I would begin with a short 720p generation, save the task ID, and only then test longer durations or reference-heavy prompts.

curl https://api.cometapi.com/v1/videos \
  -H "Authorization: Bearer $COMETAPI_KEY" \
  -F 'model="seedance-2-5"' \
  -F 'prompt="A continuous cinematic shot through a rain-lit night market."' \
  -F 'seconds="4"' \
  -F 'size="1280x720"'

A successful creation request returns an ID and an initial status such as queued. Poll the retrieval endpoint:

GET /v1/videos/{id}

Polling every 10–20 seconds is a reasonable starting point. Handle at least these terminal states:

• completed
• failed
• error

When the task completes, download or re-host the signed video_url before it expires.

For image-to-video, attach the reference image using input_reference and explain in the prompt what role the image should play.

The documented integration supports 4–30 second generation and 480p/720p output. Webhook-supported workflows may also be appropriate if polling is inconvenient.

Pricing and budgeting

The official Seedance list prices are:

Clip length	480p at $0.103/sec	720p at $0.231/sec
5 seconds	$0.515	$1.155
10 seconds	$1.03	$2.31
15 seconds	$1.545	$3.465

That puts a 10-second render at approximately $1.03 in 480p or $2.31 in 720p at those list rates.

The more useful number is cost per accepted clip. If a team needs three attempts on average before approving a result, the generation portion of the budget is roughly three times the single-render estimate. Storage, moderation, delivery, and human review are additional costs.

As of September 3, 2026, the model page lists resolution-specific default-group charges of:

• $0.412 for 5 seconds
• $0.824 for 10 seconds
• $1.24 for 15 seconds

Those estimates use the rates shown on the model page. Confirm the price displayed for your account before moving a workflow into production.

Practical use cases

Campaign concepts

Approved product, character, color, motion, and audio references can keep exploratory campaign work closer to an established visual system. The 30-second maximum also leaves room for setup, a product reveal, and a closing beat in one generation.

Previsualization

Longer continuous clips are useful for blocking camera moves, testing pacing, and exploring multi-shot sequences before committing to full production. Multi-round extension can help develop an existing sequence while retaining its direction.

Social video

Trailers, explainers, and narrative social posts often need several beats rather than a single short shot. Test vertical layouts, subject continuity, and repeatability before applying a generation template across a content calendar.

Product visualization

Reference images can anchor product form, color, and placement while the prompt controls the environment, motion, and camera. Review every output for unauthorized design changes before publishing.

A test plan I would use

There is no universal benchmark proving that Seedance 2.5 is the best choice for every production scenario. I would separate documented capabilities from what the current API exposes, then measure the actual workflow.

Test	Prompt design	What to score
Long narrative	30 seconds with three timed story beats and one visual arc	Continuity, subject drift, completed ending
Reference fidelity	One product image with explicit color, shape, and placement instructions	Identity retention, unauthorized changes, prompt adherence
Complex motion	Two subjects cross, become occluded, and reappear while the camera moves	Anatomy, physics, occlusion recovery, camera stability

Run every prompt at least three times using the same duration and size. Record:

• Success rate
• Queue time
• Render time
• Cost per accepted clip
• Reviewer score

Do not compare one attractive result from Seedance 2.5 with an average result from another model. Use identical prompts and enough repetitions to expose failure patterns.

When I would choose it

Seedance 2.5 makes the most sense when all three of these matter:

1. A longer clip
2. Controlled visual references
3. Continuity across multiple story beats

That points toward product spots, story-driven advertisements, educational sequences, previsualization, and reference-led creative development.

I would choose another route when the requirement is currently documented 1080p or 4K output, when every advertised video and audio reference type must be available through one endpoint, or when the job only needs inexpensive short drafts. I would also avoid standardizing on advanced editing features until the exact endpoint documents them.

Alternatives worth testing

Wan 3.0

Wan 3.0 is worth evaluating when high-volume drafts and lower generation cost are the priority. Its public catalog lists text-to-video, image-to-video, and video editing, with a discounted starting price of $0.04 per second as of September 3, 2026.

Compare it with Seedance 2.5 on motion quality, audio requirements, reference fidelity, and continuity—not only price.

MiniMax H3

MiniMax H3 may be a better fit when native audio, flexible reference control, and output up to 2K are more important than a 30-second Seedance generation. Its discounted starting price is listed as $0.064 per second as of September 3, 2026.

Again, measure cost per accepted clip. A lower per-second price can disappear quickly if the model requires more retries.

Bottom line

Seedance 2.5 is most compelling for longer, reference-controlled video where continuity matters more than the cheapest possible draft. Its 30-second generations, multimodal references, extension workflow, and editing controls make it a reasonable candidate for product storytelling, ads, previsualization, and other structured creative work.

The API is available through both BytePlus and CometAPI as of September 3, 2026. Before production, verify the live model identifier, confirm which parameters your endpoint exposes, and check the rate attached to your account.