Seedance 2.5: What the API Actually Ships

ByteDance's Seed team released Seedance 2.5 on July 31, 2026, and the developer API followed through Volcano Ark and BytePlus ModelArk over the next two to three weeks, with third-party routers such as EvoLink confirming billed, working generations by mid-August. That staggered rollout is worth naming up front, because most of what is circulating online treats "released" and "API live" as the same date. They were not, and a production team deciding when to actually integrate this should work from the API-confirmed dates, not the announcement date.

Here is what is actually confirmed, separated from what is still a vendor claim.

What shipped

Two changes matter more than the rest of the spec sheet.

• 30-second single-shot generation, up from 15. ByteDance frames this as the difference between generating a clip and completing a scene. Multi-round extension is supported on top of that, chaining 30-second segments while holding character and environment consistency, which is the part that actually determines whether this replaces a cut in your assembly.
• 50 reference inputs per request, split across 30 images, 10 video clips and 10 audio clips. That is more than triple Seedance 2.0's combined limit of 15. For anything with more than one or two locked characters, this is the number that decides whether the model can hold your cast and location set in a single pass.

Alongside those: clay-render referencing for spatial blocking, timestamp-level editing for targeted fixes without a full regeneration, and improved camera-perspective and green-screen handling.

Where you can actually reach it right now

Developer API access runs through Volcano Ark for China-facing accounts and BytePlus ModelArk for international accounts, with EvoLink and a handful of other routers layered on top. Creator-facing products (Jimeng AI, Doubao Pro, Dreamina) had it first. If you are trying to test this today, ModelArk is the international entry point; expect the usual enterprise-account setup rather than a self-serve key.

4K is listed as an expected capability in ByteDance's own announcement, not a confirmed specification. Independent verification through mid-August had only reached 1080p.

What is not settled

• Pricing. ByteDance has not published official 2.5 rates. Anything you see quoted as "the price" right now is a third-party router's rate or an extrapolation from Seedance 2.0's per-second cost. Do not put a number in a client SOW until you have it from your own account, not a blog post.
• Resolution. Same caveat as above. Treat "native 4K" as a claim to verify on your own account before you promise it downstream.
• Copyright exposure. Multiple studios reportedly sent cease-and-desist letters over training data and generation of recognizable IP; no lawsuit had been filed as of the reporting available. Separately, data submitted through BytePlus is subject to Chinese jurisdiction, which is a real consideration if a client's proprietary reference footage or brand assets are going through the API. This is a legal question for your own counsel, not something this article can resolve, but it is not something to skip past either.

What this means for a production stack

The reference-input jump is the genuine capability change here, not the duration bump. Going from 15 to 30 seconds is useful; going from 15 to 50 combined references is what actually changes how many locked characters and locations you can carry into one generation without a compositing workaround. If your bottleneck has been holding a cast consistent across a multi-character spot, this is the number to test against your own footage before deciding anything else about the model.

The reference-input jump is the real change

Maximum reference inputs per request

Reference inputs per request

Model	Images	Video clips	Audio clips	Total
Seedance 2.0	not broken down			15
Seedance 2.5	30	10	10	50

Everything else (the price, the resolution ceiling, the legal exposure) is unconfirmed enough that committing a client deliverable to it this month would be premature. Test it on your own account, on your own reference set, before it touches a paying job.