The short answer: On October 7, 2026, Shengshu Technology launched Vidu Q4 Preview, the first public preview of its next-generation flagship AI video model for expressive audio and video. It supports up to 15 image references and 3 audio references per generation, outputs in 2K or 4K with 10-bit color, and carries launch pricing starting at $0.014 per second. The preview is live now through Vidu’s web product and API platform at platform.vidu.com.
If you create video content — for clients, for social channels, or for product marketing — the number that matters most is the cost of iteration. A production-ready shot rarely survives its first generation. Vidu is betting that cheaper seconds plus richer reference control will change the economics of how many takes you can afford.
What Vidu Q4 Preview actually is
Vidu Q4 Preview is not a finished product launch; it is the first public preview of the next-generation flagship model from Shengshu Technology, the Chinese generative AI company behind the Vidu video model. The preview designation matters: features and pricing may shift as the company gathers feedback before a full release.
The model is built around expressive audio and video, with three headline capability areas: character performance (more nuanced expressions, clearer emotion, more natural movement), camera work (stable camera moves and cuts that follow the creator’s direction), and visual effects (improved realism, lighting, texture, and depth). It also coordinates facial expression, emotion, body movement, and voice within a single generation.
Confirmed specifications
- Reference images: up to 15 per generation
- Reference audio: up to 3 audio clips, used to guide voice consistency and emotional delivery
- Output resolution: 2K or 4K
- Color depth: 10-bit
- Availability: web product and API platform, live at launch (October 7, 2026)
- Text rendering: the company highlights stronger rendering of Chinese text, described as on par with doing your own post-production effects work
The 10-bit color detail is not marketing fluff for anyone doing client work: 10-bit footage holds up better in color grading and professional editing workflows than the 8-bit output many AI video tools produce. If you have ever watched AI-generated footage band and posterize after a grade, that is the problem 10-bit addresses.
Pricing, decoded: $0.014/sec is the headline, not the whole story
Vidu’s announcement leads with launch pricing starting at $0.014 per second of generated video. That figure positions the model aggressively: the same announcement materials note tiered pricing starting at $0.045 per second at 540P and scaling up to $0.39 per second at 4K. Read the fine print accordingly — the entry price reflects the lowest tier and resolution, and 4K generation costs substantially more.
Two pricing details worth knowing:
- Image-to-video and reference-to-video generation carries a 30% discount through November 30. Since reference-to-video is arguably Q4 Preview’s flagship workflow, the discount lands exactly where the model’s selling point is.
- For comparison, Vidu’s previous-generation Q3 model was listed at 24 credits per second at 1080p with 40 free monthly credits, so the Q4 Preview headline rate is a deliberate undercut.
To put real numbers on a real job: a 10-second clip at the launch rate costs $0.14 before resolution scaling. At the 4K rate, that same 10 seconds costs $3.90. The honest way to budget is to assume your final deliverables will sit somewhere between those two poles, and that you will pay for every failed take, too. The pricing makes iteration cheaper, but the reference-based workflows make iteration less necessary — that combination is the actual value proposition.
Why 15 image references matters more than the price
Most AI video generators accept one or two reference inputs: a starting frame, maybe an ending frame, a character sheet. Vidu Q4 Preview accepts up to 15. That is a different kind of control.
With 15 references, a creator can load a character’s face from multiple angles, the product being advertised, the wardrobe, the location, and the lighting reference — and hold them consistent across a shot. Previous reference-based tools forced creators to choose which consistency problem to solve. More references means fewer problems to trade off against each other.
Pair that with up to 3 reference audio clips guiding voice consistency and emotional delivery, and the model starts to look like a tool for a specific, high-value job: short commercial and narrative scenes where the same character must speak, emote, and stay visually consistent. That is exactly the job description Vidu gave the model: independent creators, small studios, and production teams working on narrative and commercial projects.
A practical workflow example
Consider a small studio producing a 30-second product spot. Instead of generating a single clip and hoping the product, model, and setting all match the brief, the studio could supply: product photos from three angles (3 references), the spokesperson’s reference images (3), the setting/wardrobe references (4), and lighting and mood references (2) — with headroom for more. The audio references then keep the voiceover consistent with the on-screen talent. Whether the model honors 12+ simultaneous references as faithfully as the marketing suggests is the question independent testing needs to answer, and it has not yet been answered publicly. Treat the capability as promising, not proven.
How Vidu Q4 Preview compares to the field
The AI video market in 2026 is crowded and price-sensitive, especially after Sora’s shutdown consolidated buyer attention around the surviving players. Here is where Vidu Q4 Preview’s headline numbers sit against the published rates of its main rivals, all verified from current coverage:
| Model | Representative per-second rate | Where it stands |
|---|---|---|
| Vidu Q4 Preview | From $0.014/sec (launch); $0.045/sec at 540P to $0.39/sec at 4K | Lowest headline entry price; 15 image refs is the reference-count leader |
| Kling 3.0 | ~$0.07–0.14/sec via API | Value pick for photoreal human motion; multi-shot storyboards for narrative consistency |
| Google Veo 3.1 | $0.10–0.75/sec on Google APIs (Standard ~$0.40/sec) | Consistency at the top end, 4K, deep Google ecosystem integration |
| Runway Gen-4.5 | ~$0.12/sec (12 credits/sec at $0.01/credit) | Bundled studio: editor, third-party model access including Kling |
A few honest caveats about this comparison. First, per-second list rates are not the same as cost per usable clip: a cheaper rate that needs three rerolls costs more than a pricier rate that lands on the first take. Second, credit-based subscription plans can beat pay-per-second pricing for heavy, predictable workloads. Third, Veo’s top-end quality and Runway’s bundled editing suite justify their premiums for specific professional pipelines. Vidu’s pitch is strongest for creators whose bottleneck is reference control and iteration cost — which is a real, specific segment, not the whole market.
Related coverage: if you are still comparing tool options, our roundup of free AI tools for video creation and our guide to the best AI video editors in 2026 cover the broader stack these models plug into.
Who should try it — and who should wait
Good fit
- Small studios and freelancers producing commercial spots where character and product consistency across shots is the job. The 15-reference capacity plus 10-bit 4K output targets exactly this work.
- Independent creators experimenting with narrative shorts who need affordable iteration. The launch pricing is explicitly designed for the try-one-more-shot workflow, in CEO Yihang Luo’s words: “Advanced video models should prove themselves beyond carefully selected demos.”
- Developers building video features into apps. API access at a low per-second rate is the economic foundation for programmatic generation (product demos, localized ad creative, variations).
Wait for more evidence
- Anyone needing proven character consistency at 12+ references. The headline capability has no independent verification yet — this is a preview, and preview demos are carefully selected by definition.
- High-volume production pipelines. Launch pricing and the November 30 discount window may not represent the steady-state price. Budget on the tiered list rates ($0.045–$0.39/sec), not the headline.
- Teams already deep in another ecosystem. If your pipeline runs on Veo through Google Cloud or Kling through Runway’s bundle, switching costs (workflow, prompt libraries, consistency testing) can exceed the per-second savings.
Also this week in AI models: Reka Rho-1
On October 5, 2026 — two days before the Vidu launch — US AI startup Reka released a research preview of Rho-1, a 19-billion-parameter model that takes a fundamentally different bet. Where Vidu and its rivals compete on better video pipelines, Rho-1 tries to erase the pipeline entirely: one neural network, trained from scratch, that understands and generates text, images, and video, reasons over them, and outputs robot actions — all as tokens in a single shared context window, with no tool calls and no second model.
Reka demonstrated the loop in an unedited five-turn session: the model draws a lighthouse, locates it in the scene, animates it, changes the weather in the video to a snowstorm, and explains what changed. The base model generates video at about 0.79x real time, and a distilled variant cut denoising from 99 steps to 8, producing a 5.3-second clip in about a second in Reka’s internal tests. The checkpoint trained on 320 H100 GPUs over roughly three months.
The important caveats: Rho-1 is a research preview only. There are no public weights and no public API; Reka is inviting collaborators. Robot demos run in LIBERO simulation, with no physical-robot results released. Reported video resolution is 672×384, with acknowledged drift on long videos. This is an architecture demonstration, not a product you can use today — which is exactly why it belongs in a week-in-review note rather than a review.
Why it matters anyway: if unified models like Rho-1 (and similar efforts such as NVIDIA’s Cosmos 3) mature, the “one specialist model per modality” pipeline that today’s video tools depend on could compress into a single conversation. That is a research horizon, not a buying decision — but it is the horizon Vidu’s generation of tools is racing toward.
This cluster continues with our other stories this week: Nano Banana 2.1: Google’s 4K image model, Windows hybrid intelligence from Microsoft and NVIDIA, and the challenged Anthropic AI-scientist discovery claim.
Frequently asked questions
When did Vidu Q4 Preview launch?
October 7, 2026. It is the first public preview of Shengshu Technology’s next-generation flagship AI video model, available through Vidu’s web product and API platform.
How much does Vidu Q4 Preview cost?
Launch pricing starts at $0.014 per second of video. Tiered list pricing runs from $0.045/sec at 540P up to $0.39/sec at 4K. Image-to-video and reference-to-video generation carries a 30% discount through November 30. Because this is a preview, pricing may change at full release.
How many reference images does Vidu Q4 Preview support?
Up to 15 image references per generation — the highest reference count among mainstream AI video models at launch — plus up to 3 audio references for voice consistency and emotional delivery.
What resolution and color does Vidu Q4 Preview output?
2K or 4K output with 10-bit color depth. The 10-bit color is the meaningful spec for professional editing and color-grading workflows.
Is Vidu Q4 Preview better than Veo 3.1 or Kling 3.0?
It is cheaper per second at the entry tier and accepts more reference images than either. But “better” depends on your job: Veo 3.1 leads on top-end consistency and ecosystem integration, Kling 3.0 leads on photoreal human motion and narrative storyboards, and none of Q4 Preview’s headline claims have independent verification yet. Compare cost per usable clip, not cost per second.
What is Reka Rho-1?
A 19B-parameter research-preview model released October 5, 2026, that unifies text, image, and video understanding and generation plus robot actions in a single neural network. It is not publicly available — no weights, no API — and its robot capabilities are demonstrated in simulation only.
Sources
- Vidu / PRNewswire — Vidu Launches Q4 Preview, Making Flagship AI Video Creation Accessible to Everyone (Oct 7, 2026)
- Shengshu Technology / GlobeNewswire — Shengshu Technology Launches Vidu Q4 Preview (Oct 7, 2026)
- Unite.AI — Vidu Releases Q4 Preview of Next-Generation Flagship AI Video Model
- The Decoder — Reka AI’s omni-model Rho-1 handles text, images, video, and robot control in a single model (Oct 5, 2026)
- CloudZero — Kling AI pricing in 2026: plans, credits, and API costs (competitor pricing context)
Join the discussion
Have a question or a perspective to share?