Seedance 2.5: the thirty-second clip, and what it does to a video budget
The model that already topped the video arena doubled its clip length and learned to edit a frame without regenerating the shot. Here is the verified feature list, the numbers that are still rumors, and where it lands in the toolkit.
Seedance 2.5 was announced on June 23, 2026 at ByteDance's Volcano Engine FORCE conference and shipped on July 31. It builds on Seedance 2.0, which spent early 2026 as the top-ranked video model on the Artificial Analysis Video Arena for both text-to-video and image-to-video. Monolith's own pipeline has run Seedance since the brand film on our homepage, so this one gets read closely.
What is confirmed
| Capability | Seedance 2.0 | Seedance 2.5 |
|---|---|---|
| Single-pass clip length | 15 seconds | 30 seconds, with scene and tempo changes inside the clip |
| Reference inputs | 12 | Up to 50: 30 images, 10 videos, 10 audio |
| Editing | Regenerate the clip | Timestamp-level edits, local redraw, green-screen replacement |
| Lip-sync | Multi-language | 20 languages with matched lip-sync |
| Extension | Chain generations | Multi-round extension in one workflow |
| 3D references | Not offered | Clay-render referencing for consistent geometry |
Two of those rows are the whole story. Thirty seconds in a single pass means a product spot, a testimonial-style piece, or a chapter of a brand film comes out as one continuous performance instead of four stitched shots pretending to be one. And timestamp-level editing means a wrong hand gesture at second nineteen gets fixed at second nineteen, not by rolling the dice on a whole new take. Anyone who has burned an afternoon regenerating a fifteen-second clip to fix a logo will understand what that does to a day rate.
What is not confirmed, and why the copy says so
Resolution. Several third-party hosts advertise 4K, and one round of coverage credits Seedance 2.0 with native 4K and 10-bit output, but ByteDance has not published official resolution figures for 2.5. Pricing for 2.5 is described as token-based resource packs without public per-second rates; hosts list Seedance 2.0 at roughly 15 to 80 cents per second depending on resolution. And there are no independent benchmark scores for 2.5 yet. Digital Applied put it bluntly: any Seedance 2.5 score circulating now is invented. Monolith agrees, which is why this post has an arena chart for 2.0 and none for 2.5.
Where it lands in a working toolkit
Monolith does not run one video model; it runs whichever is best for the shot. Seedance is the workhorse for character consistency and continuous performance, Veo for certain photoreal and dialogue work, Kling and Runway where their motion or control tools fit the brief. What 2.5 changes is the ratio: more of a thirty-second piece can now come from one model in one pass, with edits instead of re-rolls, and the fifty-reference limit means a whole character sheet, a location set, and a soundtrack can ride along on every generation. That is the discipline behind the 70M+ views the team produced for the LitVM channel, now with less friction.
- Product and service spots: one thirty-second continuous take with the real product in the reference set.
- Character brands: a mascot, a recurring host, or a founder stand-in kept on model across a month of daily posts.
- Localized campaigns: one performance, twenty languages of matched lip-sync, no reshoot.
- Fix-in-place revisions: the client note at second nineteen becomes a timestamp edit, not a new round.
AI video, music, and sound from Monolith's production pipeline: sprints from $1.5k, retainers for brands that need a daily cadence, all on the current best models with a creative director on every cut.
For the proof, watch Hard Money Livin', a full rap music video produced with no camera and no shoot day.
For what those per-second rates do to a whole production budget, and why direction is now most of the invoice, read the cost breakdown.
What is new in Seedance 2.5?
Does Seedance 2.5 output 4K?
What does Seedance cost per second?
Seedance 2.5 doubles the clip and fixes frames in place. Treat the 4K claims as unverified until ByteDance says otherwise, and treat the thirty-second continuous take as the new unit of brand video.
This is the thinking behind a service Monolith runs every week.
The prompt library every small team should have
25 copy-paste prompts for leads, marketing, operations, hiring, and admin. The ones we actually give clients, as a designed PDF. Free, in exchange for an email.