Seedance 2.5 Officially Launches: One-Take 30s AI Video

Seedance 2.5 officially launched today, July 31, 2026. ByteDance’s Seed team published the launch post today — “Today, we are officially launching Seedance 2.5, the new-generation video creation model” — and the model went live on two named consumer surfaces, Jimeng Web and Doubao Pro, the same day.

This is the moment the June announcement was pointing at. We covered the enterprise-beta reveal in our June 23 Seedance 2.5 announcement post — this article covers only what changed today: the June feature set moving from announced to shipping, plus the genuinely new material in the launch post — multi-round extension from 30 seconds to several minutes, clay-render referencing, timestamp-level and camera-perspective editing, and an API timeline that ByteDance’s own launch post quietly corrects.

One more thing worth saying up front: as of publication, Western tech press has not yet covered today’s launch. Everything below is sourced from ByteDance’s official launch post, with pre-launch reporting and independent leaderboard data used only where dated and attributed.

ByteDance names two surfaces on day one: Jimeng Web (Video Generation → Select Seedance 2.5) and Doubao Pro (same path). The launch post adds “and other platforms” without naming them, so treat two as the confirmed floor rather than the total. Both named surfaces are ByteDance’s China-market creative apps, which makes this a China-first consumer rollout — 38 days after the June announcement.

ByteDance describes 2.5 as extending “the unified multimodal audio-video joint-generation architecture of Seedance 2.0” toward both foundational generation and reference-based generation, and frames the release as “elevating video generation from clip-level outputs to comprehensive creative workflows.” Strip the vendor language and the claim is concrete enough: longer single takes, more reference material per job, and editing operations that work on video the model has already generated.

The headline capability: “Seedance 2.5 can generate high-quality, 30-second audio-video clips in a single pass.” That doubles Seedance 2.0’s native 15-second single-pass ceiling — beyond which 2.0-era workflows required stitching clips together in post-production, with all the continuity seams that implies.

The more consequential mechanic is multi-round extension. Users can “smoothly append subsequent shots to existing video outputs,” with ByteDance claiming consistency of main characters, environments, and narrative pacing across the appended shots — producing several minutes of content from one workflow. If the consistency claim holds up outside vendor demos, that moves the practical unit of AI video from the clip to the sequence.

Seedance 2.5 accepts up to 30 images, 10 video clips, and 10 audio clips — 50 combined reference inputs — in a single pass. That is the same 50-input ceiling ByteDance announced in June, now confirmed at launch. The 2.0 baseline it is measured against is less settled: TechTimes reports roughly 15 combined assets (9 images, 3 video clips, 3 audio files), while the figure circulating at the June announcement was 12 — so call the increase roughly three- to fourfold rather than a precise multiple. For production teams, that is the difference between referencing a character and referencing a whole art-directed world: cast, wardrobe, location plates, motion references, and a temp score in one job.

Clay-render referencing

The genuinely new input mode is clay-render referencing: users block out a scene’s spatial structure, character poses, motion paths, and camera angles using textureless 3D models — the grey “clay renders” familiar from previsualization — and the model generates finished video matching that blocking. ByteDance also says the clay render’s spatial data drives lighting: the model generates “realistic lighting effects that follow physical laws, such as light source direction, color temperature, intensity, and shadow projection.”

That is a previs-to-final pipeline in one step, and it is aimed squarely at people who already work shot-by-shot. If you are new to the model family, our Seedance 2.0 guide covers the baseline referencing workflow this builds on.

June’s announcement already promised editing — local re-draw that changes one element of a frame without disturbing the rest. Today’s launch post documents a four-mode suite instead, and the two modes that go furthest past that June claim are timestamp-level and camera-perspective editing: they change when and from where a shot happens, not just what sits inside the frame. All four operate on video the model has already produced, rather than forcing a full regeneration.

The pattern across all four is the same: preserve what works, change one variable. That is how human editors think, and it has been the missing primitive in AI video — where the answer to “can you just move the camera?” has historically been a fresh generation and a fresh continuity lottery. It is worth stressing these are vendor-described capabilities as of today; independent hands-on testing outside China-market apps does not exist yet.

Most coverage lists 2.5’s features in isolation. The table below diffs today’s confirmed launch facts against ByteDance’s own stated 2.0 baseline — what shipped, what is new, and what remains unpublished on both sides.

Sources: ByteDance Seed launch post (Jul 31, 2026) for all 2.5 cells; TechTimes (Jul 16, 2026) for the 2.0 baseline figures; Artificial Analysis leaderboard checked Jul 31, 2026.

Here is the discrepancy worth your attention if you are planning a build. On July 16, TechTimes reported that ByteDance had “opened public API access to Seedance 2.5” through BytePlus — framing the rollout as complete. Fifteen days later, ByteDance’s own launch post says API access is coming soon via BytePlus ModelArk, and BytePlus’s own ModelArk model-list documentation, checked today, carries tutorials and prompt guides for Seedance 2.0 only — no 2.5 entry. The honest reading: whatever opened on July 16 was a narrower access tier, and full ModelArk availability is still pending as of today’s consumer launch.

Why it matters: teams outside China cannot build on Seedance 2.5 today, full stop. The consumer surfaces are China-market apps, and the documented API path still serves 2.0. Any agency roadmap that penciled in “Seedance 2.5 API, live since mid-July” — and we have seen that assumption circulating — is running ahead of ByteDance’s own record. Watch the ModelArk model list, not the press cycle.

Today’s launch post is notably restrained. It contains no price, no resolution specification, and no benchmark or arena claim — confirmed by a direct read of the full post. That restraint matters because the rumor mill has been less careful. The 4K figure that circulated pre-launch was, per TechTimes’ own July 16 reporting, “an expected capability, not a confirmed specification” — ByteDance has never officially stated 4K for 2.5, and today’s primary makes no resolution claim at all.

The “#1 video model” framing deserves the same scrutiny. Claims of Seedance topping the charts rest on weighted composite scores, not head-to-head preference Elo. On Artificial Analysis’s Elo-based text-to-video arena — blind human-preference votes — the shipping Seedance 2.0 sits third, and Seedance 2.5 has no entry yet at all.

The competitive backdrop shifted again today: MiniMax launched H3 the same day — our MiniMax H3 launch coverage breaks down Seedance’s nearest same-day competitor, which already sits 17 Elo points above shipping Seedance 2.0 on the same board. For the wider field, our Flux 3 vs Seedance 2.5 vs Gemini Omni comparison maps how the contenders stack up.

On cost: no Seedance 2.5 pricing exists anywhere as of today. The widely-cited figure — roughly $9 per minute for Seedance video-with-audio versus roughly $24 per minute for Google Veo on Vertex AI, per Artificial Analysis — is a third-party estimate from the 2.0 era that predates this launch and has not been confirmed to apply to 2.5. We dig into what the pricing chaos means for ad budgets in our AI video price-war analysis, published alongside this post.

One paragraph of context, because it shadows every commercial decision here. In late February 2026, the Motion Picture Association sent ByteDance a cease-and-desist letter over Seedance 2.0 — its first ever to a major generative-AI company — alleging “systemic infringement,” following individual letters from Disney, Warner Bros. Discovery, Paramount Skydance, Netflix, and Sony Pictures (per the Hollywood Reporter). Separately, SAG-AFTRA condemned what it called “blatant infringement” of its members’ voices and likenesses, as TechTimes reported on July 3. As of mid-July, no U.S. studio had filed a federal lawsuit — reportedly because serving legal process on a Beijing-headquartered company under the Hague Service Convention takes an estimated 18 to 24 months. ByteDance, for its part, previewed a copyright-licensing platform at the June 23 FORCE conference with filmmaker Stephen Chow as an initial partner. Nothing in today’s launch post addresses any of this.

ByteDance itself concedes “there is still room for improvement, particularly regarding the physical plausibility of complex motions and the stability of scenes involving interactions among multiple subjects” — a limitation disclosure worth taking at face value before anyone promises a client multi-minute one-take spots. Here is how we would route the decision by team type.

The trend worth reading here goes beyond one model: with 2.5, ByteDance is competing on workflow depth — editing, referencing, extension — while rivals compete on output quality scores. That suggests the next competitive phase in AI video is about who owns the revision loop, not who wins the arena vote, because revision is where production time and budget actually go. Looking forward, we expect the ModelArk listing to land with pricing attached, and that pricing — not today’s feature list — is what will determine whether 2.5 displaces incumbent tools in ad-creative pipelines. We built an ad-creative pipeline around Seedance 2.0 earlier this year; the architecture transfers, and teams planning similar builds can pressure-test the approach with our AI transformation engagements or fold the output into managed paid media programs once the API materializes.