AI video generation just crossed a threshold that changes what we tell clients to build first this quarter. Two flagship models shipped eight days apart this summer, and both do things that were not possible six months ago: locked character consistency across scenes, native audio, and single-take clips long enough to carry a real ad or a full property walkthrough. We have spent the past two weeks running both through our own content pipeline, and here is what actually holds up.
Two flagship video models shipped days apart
ByteDance's Seedance 2.5 launched on Jimeng and Doubao Pro on July 31, 2026. You can test it now through Dreamina, the ByteDance platform that also hosts Seedance 2.0 and the Seedream image models, or through the Chinese consumer app at Jimeng. The headline spec is a single-take clip up to 30 seconds long, built from up to 50 reference files in one pass: images, video clips, and audio combined. That reference count matters more than the runtime. Feed the model a locked set of brand assets, a product shot, a voice sample, a location plate, and the output stays consistent with what your brand actually looks like instead of drifting into a generic AI face.
Black Forest Labs answered eight days earlier with Flux 3, its first video model and the first model in its lineup trained jointly across image, video, and audio. Flux 3 generates clips up to 20 seconds with native audio built in: dialogue, sound effects, and ambience, at no extra cost. Pricing on the official page runs from around six cents a second for draft-quality HD text-to-video up to twenty-nine cents a second for standard FHD, and a five-second sample in standard HD costs about eighty-five cents. The open-weight version, Flux 3 Dev, is planned for later this year with no date confirmed yet.
The same week, Alibaba shipped its own flagship, Qwen3.8-Max, a 2.4 trillion parameter model. We are not routing production work to it, but the pattern is the point: three labs shipped frontier releases inside eight days. If your content stack still assumes one model forever, that assumption is already out of date.
Why this jump reads differently than the last twenty "better video model" releases
Video generation has improved every few months for two years. Most of those releases were incremental: slightly sharper textures, slightly longer clips, the same tell-tale AI look on human faces and hands. Seedance 2.5 and Flux 3 close two gaps that actually change what you can publish. First, the gap on human faces has narrowed enough that a casual scroll will not catch it, especially in close-ups under normal lighting. Second, both models now hold a character or a product steady across multiple shots inside one generation, instead of resetting the look every clip. That second point is the one that matters for a business, not a hobbyist: it means brand-consistent video at volume, not a pile of one-off clips you have to cherry-pick from.
What each model actually offers today
| Seedance 2.5 | Flux 3 Video | |
|---|---|---|
| Max single clip | 30 seconds | 20 seconds |
| Reference inputs | Up to 50 files (image, video, audio) | Image and video-to-video conditioning |
| Native audio | Included | Included at no extra cost |
| Price signal | Consumer access via Dreamina; API pricing not yet public | $0.06 to $0.29 per second by quality tier |
| Access today | Dreamina, Jimeng, Doubao Pro | Black Forest Labs dashboard and API, gated early access |
| Open weights | Not announced | Flux 3 Dev planned, no date set |
The AI slop problem is now a trust problem for your brand
Better tools do not just produce better video. They also produce more junk, faster. As the line between real and generated footage blurs, the risk shifts from "can we make this" to "should we publish this without saying so." A viewer who cannot tell your product demo from a generated clip will eventually stop trusting either one, and that costs more than the subscription fee. We treat this as a production discipline problem, not a model problem. Every clip that ships gets three checks before it goes live. A human watches it full length, the brand assets used as reference get logged, and anything customer-facing that claims to show a real result gets flagged and disclosed where it should be. The tool got better. The review gate has to get stricter at the same pace, or the tool becomes a liability instead of a lever.
What we are testing this quarter
We run a 40-person real estate group alongside the agency, so property video is the first place we are testing this. A locked reference set, floor plan, five to ten still photos, one voice sample, turns into multiple listing walkthroughs and social cuts from a single generation pass, instead of a separate shoot per property. For client work, the same pattern applies to ad creative: one reference library per brand, dozens of variations generated against it, and a human review queue before anything gets a spend behind it.
- Reference-locked generation for every recurring content type: listings, product demos, testimonial-style ads.
- A published disclosure policy for any customer-facing AI-generated video.
- Cost-per-video tracking instead of cost-per-subscription, since Flux 3's per-second pricing makes the real unit economics visible for the first time.
- A watch list on Flux 3 Dev's open-weight release, since self-hosted generation changes the calculus for anyone producing video at real volume.
How this maps to the systems we build for clients
This is the same pattern behind every system we build: intake the source material once, automate the repeatable output, and put a human at the one checkpoint that actually needs judgment. For video, that pipeline looks like a reference library feeding a generation queue feeding a review step feeding the publish channel, running on its own instead of waiting on a designer's calendar. See how we structure this kind of build under automation, or look at how it plays out end to end in our case studies.
If your team is already spending real hours per week on manual content production, that is exactly the kind of leak our revenue leak heatmap is built to surface. A lot of operators are sitting on ten or more hours a week of admin and production work that a system could carry instead. The starting point for any of this is our assessment, a flat €999 that gets credited in full toward the build. Most first systems go live in days to weeks, not quarters, and clients own 100% of the code and files that come out of it. You can see the full product set, including AI Concierge, AI Operations Agent, and Second Brain, on our products page, and we post the rest of our testing notes on the blog as we run them.
What to build this quarter
- Pick one recurring, high-volume content type (ads, listings, demos) and pilot reference-locked generation on it before touching everything else.
- Build the reference library first: five to ten locked assets per brand or property beats one clever prompt.
- Put a named human on the review gate before anything customer-facing ships, no exceptions.
- Track cost per finished video, not the subscription price, so you can compare Seedance, Flux 3, and whatever ships next on the same basis.
- Write down your disclosure policy now, before a customer asks whether a video is real.
Common questions
Is Seedance 2.5 available outside China?
Consumer access runs through Dreamina and Jimeng today, and Doubao Pro also carries it. Volcengine's Ark API listed the model as coming soon as of its official launch post, so wider developer access is close but not fully open yet.
How much does Flux 3 video generation cost?
Black Forest Labs' official pricing runs from about six cents a second for draft HD text-to-video up to twenty-nine cents a second for standard FHD, with video-to-video generation priced higher. A five-second standard HD sample runs about eighty-five cents.
Should we replace our video agency with AI video generation?
Not outright. Reference-locked generation is strong for volume: ad variations, listing walkthroughs, product cuts. Hero brand films and anything that needs real human performance still belong with a production team. The win is using generation to clear the long tail so your agency budget goes to the work that actually needs it.
What is the fastest way to start testing this?
Start with one content type and one locked reference library, not a full rebuild of your content operation. That is the same scoping we run in our assessment before any build starts.