How to Access Seedance 2.5: Key, Request, First Render

ByteDance announced Seedance 2.5 on 31 July 2026, and hosts picked it up at different times afterwards. That staggered rollout is why how to access Seedance 2.5 is a harder question than it looks: plenty of surfaces advertise it while still serving the older model underneath. Getting a first render means confirming the version, holding a key, and sending a request whose defaults you have actually read.

First, confirm the version is real

Two checks settle it, and neither costs anything.

Ask for 20 seconds. Seedance 2.5 runs from 4 to 30 seconds in one generation. Seedance 2.0 stops at 15. A 20-second request rejected on the duration field means you are on 2.0. A validation rejection happens before a task is created, so nothing is charged.

Read the model identifier. The released 2.5 build is served under a string ending in -face, written in full as doubao-seedance-2.5-face. Code samples published before the release quote an identifier without that suffix, and it no longer resolves. If a platform's documentation still shows the old string, its integration predates the release.

A third check that is not conclusive but is suggestive: if the interface offers a 4K tier, it is describing pre-release marketing. The model tops out at 1080p, with 720p and 480p below it, on every host.

Get a key and a place to put files

An API key from the platform dashboard, sent as an Authorization: Bearer header, is the whole of authentication. Treat it as a spending credential and keep it on a server.

If you intend to use reference material, you also need somewhere to host it. Media inputs are public HTTP or HTTPS URLs only. Data URIs are rejected across the platform, so local files have to be uploaded before they can be referenced.

How to access Seedance 2.5 with a first request

The call is a POST to the platform's video generation endpoint. It returns a task id, not a video. You then poll a task endpoint until the status is terminal. Polling is free.

The minimum useful body is a model string, a prompt and a duration. The defaults underneath matter more than most people expect:

The duration default of 5 seconds is a sensible first render. The -1 value is the one to avoid early, because the reservation is taken at the ceiling rather than at whatever the model eventually produces.

Prompts run to 20000 characters. A prompt is required for text to video and becomes optional once reference media is attached.

The reference media door

This is the capability that separates 2.5 from the older version, and it is reachable on the first request. One job accepts up to 30 reference images, 10 reference video clips and 10 audio tracks.

The pricing asymmetry is the thing to internalise. Reference images and reference audio add nothing to the bill. Reference video does, and it does so twice over. A source clip is charged for its own length on top of the output, and any request containing video carries a billing floor of five thirds of the output duration. A two-second source with a five-second output bills nine seconds rather than seven. Each reference clip runs 2 to 30 seconds and the combined reference video cannot exceed 30 seconds.

So a first render conditioned on stills is close to free of surprises. A first render conditioned on footage is not, and should be small.

What the first clip costs

Rates are per second of finished video, set by resolution and by whether the request carried a reference clip:

A five-second 480p clip is about $0.59. A ten-second 1080p clip is about $4.62. The distance between those two numbers is the argument for making the first attempt short and small, confirming the prompt does what you meant, and only then raising resolution and length.

Generation takes two to five minutes. Failed jobs are refunded in full, refusals included.

When the first call fails

A safety refusal returns error 80006. It does not name the stage that fired, so the prompt, any reference image and the generated output are all candidates, and diagnosis is manual: reduce to a neutral prompt, confirm it passes, add elements back one at a time.

The content_filter parameter defaults to true. Setting it to false routes the request to a less restrictive channel and does not disable moderation. Below that sit model-level blocks on named real celebrities, third-party intellectual property and illegal content, which no host setting removes.

Everything else in the first-call failure list is validation: a duration outside 4 to 30, a retired model string, a data: URL, a reference clip over the length limit. None of those create a task, so none of them cost anything.

Frequently asked questions

Do I need a different account to access Seedance 2.5 instead of 2.0? No. Platforms serving both expose them as separate model identifiers on one account and one key. Moving between versions is usually a single string change.

How do I know a platform is really running 2.5? Request 20 seconds. Success means 2.5, rejection on the duration field means 2.0. Then check that the model identifier ends in -face.

Can I access Seedance 2.5 without an API key? Consumer apps and bundled editor integrations reach the model without one, at the cost of parameter access and per-second billing. You get presets and a credit balance instead.

Is 4K available on any plan? No. The released model caps at 1080p. The 4K figure comes from pre-release coverage that was never delivered.

The short version

How to access Seedance 2.5 is four steps. Verify the version with a 20-second request and the -face identifier. Take a key and keep it server side. Send a short first job, five seconds at 480p, leaving the defaults alone except duration and resolution. Poll for two to five minutes and download the result. Reference images and audio are free to attach, reference video is billed for its own length with a five-thirds floor, and anything that fails is refunded.