🔍 Seedance 2.5 Specs: What ByteDance Actually Confirmed

We corrected our own Seedance 2.5 coverage this week. An earlier version of our guide led with "native 4K output" — a figure we carried over from June pre-launch reporting. ByteDance's actual launch announcement publishes no resolution at all. If we got that wrong reading the same sources everyone else did, the pages that never went back and checked are still wrong.

So here is the full ledger.

Update, August 7: the ledger moved. Higgsfield shipped seedance_2_5 in its live model catalog — the first real third-party API surface for this model — and we verified its parameters against that catalog the same day. ByteDance's own ModelArk API is still "coming soon", so what follows first is the shipped Higgsfield deployment, then the ByteDance-side ledger, most of which still stands.

Shipped: what the Higgsfield deployment actually exposes (August 7)#

Everything in this table was verified against Higgsfield's live model catalog on August 7. It documents Higgsfield's deployment, not a ByteDance publication — where the two disagree, we say so.

Three rows deserve flags.

Resolution settles our argument. This page has said since August 1 that any Seedance 2.5 resolution figure is invented. Now a shipped implementation exists, and it exposes 480p and 720p — while Higgsfield's own marketing page for the same model headlines "up to native 4K". ByteDance still publishes nothing. The 4K number remains marketing on every surface that carries it.

No start/end frames means keyframe pipelines stay on 2.0. Seedance 2.0 keeps its start_image/end_image roles; the 2.5 deployment has none. If your workflow pins a first or last frame, that row decides for you — the full split is in Seedance 2.5 vs 2.0.

The reference cap is a claim, not a parameter. Higgsfield says "up to 50 multimodal inputs" — consistent with ByteDance's 30 + 10 + 10 split — but the API publishes no enforced cap, so it stays filed as a vendor claim on both sides.

Two gaps survive the ship: no frame rate published anywhere, and no credit prices — 2.5 is credit-billed on Higgsfield and not eligible under its unlimited plans, and what a generation actually costs is unpublished. Parameter-by-parameter usage notes are in our Higgsfield settings guide.

By the numbers — confirmed#

Not published — treat any figure you see as invented#

These are the ByteDance-side unknowns. The Higgsfield deployment answers a couple of them for its own surface — flagged inline — without ByteDance having published anything.

• Resolution. No ceiling stated. Not 4K, not 1080p, nothing. (Higgsfield's shipped implementation exposes 480p/720p — a cap on its deployment, still not a ByteDance spec. The 4K headline remains marketing.)
• Frame rate. Not stated. Still not, anywhere — Higgsfield's deployment publishes no fps either.
• Pricing. No rate card, no per-second figure, no credit cost. Higgsfield has published one structural fact — video_edit bills by the reference video's duration — but no credit prices.
• Official API model IDs. BytePlus ModelArk's model list still shows Seedance 2.0, 1.5-pro and 1.0-pro. There is no 2.5 entry. (seedance_2_5 is Higgsfield's ID for its own deployment, not a ByteDance-published one.)
• Render times and latency. Nothing published.
• Error codes. Nothing published for 2.5.
• Independent benchmarks. None exist. The ~20% prompt-adherence figure is ByteDance's own, and CineD explicitly flags the absence of third-party testing.

That is a long list of unknowns for a model with hundreds of pages already ranking for its specs.

Four fabrications currently circulating#

"Native 4K." The most widespread. It comes from June reporting on the pre-launch beta and was never in the shipped announcement. We repeated it ourselves and have corrected it.

"Identity-Lock." A named feature appearing across several tool blogs and exact-match domains. It is in no ByteDance material. Consistency in 2.5 comes from the 50-file reference budget and from extension rounds that carry characters forward — not from a toggle you can enable.

"Reference strength 70–80%." A percentage slider that does not exist in any published interface description. Neither does a user-exposed seed lock.

Fabricated drift benchmarks. These matter because they shape how people spend the 50-slot reference budget. One widely-copied claim states that cutting from six references to two reduced drift by around 60%. Another describes hands-on testing where identity "drifts by the third or fourth shot." Both appear on sites publishing about a model released the day before, with no methodology. Nobody had run controlled drift tests on Seedance 2.5 in that window.

The corpus even contradicts itself: one page insists 2.5 "is not publicly available yet as of July 23, 2026, and remains in closed enterprise beta" while others describe hands-on results.

How to check a Seedance 2.5 claim yourself#

1. Does ByteDance say it? Start at ByteDance Seed's launch post. If a spec is not there, it is not confirmed.
2. Does BytePlus list a model ID? If the API is being described but ModelArk has no 2.5 entry, the writer is describing a gateway's own convention. The one exception since August 7: Higgsfield's seedance_2_5 is a live deployment you can verify in its catalog — a real, testable surface, though it documents Higgsfield's implementation rather than ByteDance's.
3. Is the number attached to a version? Most concrete Seedance facts in circulation — pricing, moderation behaviour, error codes, render times — are genuine 2.0 facts relabelled as 2.5.
4. Could they have tested it? A drift benchmark published within 48 hours of a consumer-app launch, with no methodology, was not run.

What this means for planning#

You can plan around the confirmed set. Thirty seconds in one pass, fifty reference files, and a real editing suite are substantial and well-documented capabilities — enough to decide whether 2.5 changes your workflow.

You cannot plan a budget or a delivery spec. Without a published resolution you do not know if it meets a client's master requirement, and without pricing you cannot model cost per finished minute. If a project depends on either, stay on Seedance 2.0, whose numbers are known, until ByteDance publishes.

August 7 softened half of that. On Higgsfield you now have a delivery ceiling — 720p, with upscaling for anything above it — and a real API surface to prototype against. The budget half stands: no credit prices are published, so cost per finished minute is still unmodelable, on Higgsfield or anywhere else.

For what H3 does differently — and it publishes a complete API reference, which is the sharpest contrast between the two July 31 launches — see H3 vs Seedance 2.5.

How we assessed this#

Confirmed rows come from ByteDance Seed's own launch post and its model page, cross-checked against CineD, the one trade publication covering the launch rather than reselling access. The "not published" list was verified by absence: we checked the announcement and BytePlus ModelArk's model list for each item and found nothing. The shipped-parameter table was verified first-hand against Higgsfield's live model catalog on August 7 — the first rows on this page we could check against a running surface rather than an announcement.

The fabrication examples are real claims from pages currently ranking. We have not named the sites, because the point is the pattern rather than the offender, and several are churning the same syndicated text.

Last verified: August 7, 2026 — the Higgsfield parameter table against the live model catalog, and ModelArk's model list re-checked (still no 2.5 entry). This page is only useful while it is current. When ByteDance publishes a resolution spec or a rate card, the rows move from the second table to the first and we will update them here. The August 7 update is that mechanism working — rows moved because a platform shipped, not because ByteDance published.

What didn't make the cut#

Our own measurements. We do not have Jimeng AI or Doubao Pro access, so every claim here is documentary. The August 7 parameter table is the one exception: read first-hand from Higgsfield's live catalog — verification of the surface, still not measurement of its output. We would rather publish a short verified list than a long plausible one — which is exactly the failure mode this page exists to document.