Seedance 2.5 API 'coming soon' on ModelArk: what to call today

ByteDance Seed's launch post for Seedance 2.5 says the model is rolling out on Jimeng AI and Doubao Pro, with API access "coming soon via BytePlus ModelArk" (read 2026-10-05). Sume already lists seedance-2.5 in its Video Router catalog, so you can call it today with model: "seedance-2.5".

The catalog row lists 4-30 seconds at 480p, 720p and 1080p, and a 30-second 720p 16:9 clip costs $17.34 on Sume.

What the vendor post says and does not say

The Seed post lists up to 30 seconds per generation with multiple rounds of extension, up to 30 images, 10 video clips and 10 audio clips per pass, timestamp-level editing, green-screen background replacement, camera perspective editing and reference-based editing. It gives no price, no resolution list and no API schema; those were not in the post I read.

So a team that wants to build before ModelArk ships has two jobs: pick an access path that exists now, and avoid code that ties itself to one vendor's field names.

What Sume exposes today

Sume's public docs describe the same async flow for every video model: submit to POST /v1/videos, poll the polling URL until the status is completed, then download from the content URL. Seedance 2.5 is the longest-running row in the catalog at 30 seconds, alongside Wan 3.0 (2-30 s). The narrower items are below.

Keep a later switch cheap

• Put the model id in config, not in code: SEEDANCE_MODEL=seedance-2.5. Sume uses bare catalog ids, with no provider-org prefix.
• Wrap submit and poll in one function. Sume's async job shape (id, polling URL, status, content URL) is the only thing the rest of your app should see.
• Read capabilities from the catalog before you submit, rather than hard-coding 30 seconds or a resolution list. The docs say limits are different for each model.
• Keep the price estimate separate from the call. Sume bills the provider list price x 1.25; a different access path will have a different price, and your margin math should be one function.

Read the catalog row

The catalog endpoint returns the supported resolutions, durations and reference types. Run it before you pin an id.

curl "https://api.sume.com/v1/videos/models" \
  -H "Authorization: Bearer $SUME_API_KEY"

If you need editing today

The Seed post lists editing features that Sume does not document for Seedance 2.5, so do not expect a timestamp or green-screen field on seedance-2.5. Sume documents a different route for editing an existing clip: gemini-omni-flash-1.1 accepts a video_url as an edit source, with the prompt giving the instruction. That is a separate model with its own price and limits, so test it on a short clip before you build a workflow around it.

For a clip that is wrong only at the end, regenerating with a tighter prompt is usually simpler than editing, and at 480p it costs little. Decide by cost: a 480p retry of a 10-second Seedance 2.5 clip is $2.69.

Is waiting for ModelArk worth it?

Wait only if you need something the Seed post announced that Sume does not document, such as timestamp-level editing or the larger reference counts. For plain text-to-video, first-frame and reference generation up to 30 seconds, there is nothing to wait for. If you do wait, you lose the draft-and-test cycle you could be running now: at 480p a 30-second Seedance 2.5 clip is $8.07, which is cheap enough to learn what the model does with your prompts before the vendor API lands.

This post makes no claim about when ModelArk access will open or what it will cost; the vendor has not said in the post I read.