AI Video Generation Shifts East as Kling 4.0 and Wan 3.0 Lead

Chinese labs and multimodal models now lead AI video

In the past few weeks, the AI video market has become very different from the one people pictured a year ago. Kuaishou says Kling 4.0 launched on October 15. It generates up to 30 seconds of video in a single pass and accepts as many as 15 reference assets and 10 keyframes14. Alibaba's Wan 3.0 now ranks first on the Artificial Analysis text-to-video leaderboard, both with audio and without11. OpenAI has left consumer video generation entirely. It closed the Sora app on April 26 and shut down the Sora API on September 24410.

This article argues that 2026 is the year AI video stopped being a showcase for one American lab and became a crowded, fast-moving product category. Chinese labs are setting the technical pace. Native audio and longer takes are now standard features. And many users no longer choose a model directly; they choose a platform that offers several.

What happened: an October full of launches

Kling 4.0 was the biggest release of the month. Kuaishou names three main improvements. Large-scale motion is steadier during fast action and complex camera moves. Clips can run 30 seconds natively, which allows long takes. And a feature called Omni Reference lets creators combine many reference assets and keyframes to guide a sequence14. The company also says video extension will soon reach two minutes. Kling 4.0 now connects to Kling's MCP and CLI tools, so AI agents can drive the model, and API access for enterprises and developers is coming14.

Kling 4.0 spread to other platforms quickly. One aggregator said it had the model by October 243. Runway now documents a Kling 4 API route that produces 3 to 30 seconds of 720p or 1080p video with audio7.

Other releases came close together. VideoGen launched Model 1 on October 20 and says it can produce videos up to 10 minutes long from text or image prompts. The company backed the launch with $1 million in creator credits and a claim that it will cost about half as much as comparable models12. Kandinsky Lab released the code and weights for Kandinsky 6.0 Video on October 6 under the MIT license. It comes in a 29B Pro size and a 3B Lite size, and both generate short clips with synchronized audio11. On October 30, The Biological Computing Co. launched "mischief 1." It is built on the open Wan 2.2 model, and the company says it runs five times faster than that base model. It is priced as the cheapest image-to-video option on fal and OpenRouter18.

Why it matters: Sora's exit and who replaced it

OpenAI's retreat from video is the key context for everything else this year. OpenAI announced Sora's discontinuation on March 24, 2026, and closed the product in two stages: the app in April and the API in September8. Industry guides now treat Google's Veo 3.1, Kling, and Runway as the main successors810. One comparison says the shutdown made "which model" and "which platform" two separate questions, because the same model can now be reached through many storefronts9.

Google did something different from what many expected. Many people predicted a "Veo 4." Google instead announced Gemini Omni at I/O 2026. Omni accepts text, image, audio, and video as input and returns video with synchronized sound, and it has replaced Veo inside the Gemini app4. Gemini Omni Flash entered public preview on June 30 and reached general availability as version 1.1 on August 2717. Google's developer documentation now recommends it as the default for video generation117. Veo 3.1 is still available. It has a free consumer allowance and a cheap Lite API tier17. But the older Gemini API preview endpoints for Veo 3.1 were shut down on October 227.

This is the clearest sign of the multimodal trend. Google is treating video generation as one output of a general model that handles any input and any output, not as its own product line. Black Forest Labs is doing the same with FLUX 3. That model family uses one architecture for images, video, audio, and action prediction, and it generates up to 20 seconds of video with synced sound19.

The East-West balance

Across the coverage, the leaderboards tell a consistent story: Chinese developers hold most of the top spots. Wan 3.0 reached general availability on Alibaba Cloud in August and leads both text-to-video boards. ByteDance's Seedance 2.5 ranks third on the with-audio board. MiniMax's H3 is the strongest open-weight model on the leaderboard11. Alibaba also produced HappyHorse. It appeared anonymously on Artificial Analysis in April and went straight to first place before Alibaba revealed that it was the developer17.

Runway's position is harder to read, and here the reporting disagrees. Runway says Gen-4.5 scored 1,247 Elo and ranked first when it launched in December 20256. One analysis now places it 23rd on the no-audio board as a model without native sound2. Another says it does not appear on the live leaderboards at all6. Comparisons also disagree on whether Gen-4.5 generates audio. Some tables say yes1, while other reviewers call it silent or say audio is not listed39.

The likely explanation is that Runway has changed what kind of company it is. Its own enterprise messaging stresses day-zero access to outside models such as Seedance, Kling, and Veo, alongside its own models9. Its Runway Dev platform puts one API over its in-house models plus third-party models, including Seedance, GPT Image 2, and ElevenLabs17. Its newest major releases, the Aleph 2.0 editing model and Edit Studio, focus on editing existing footage rather than on raw generation4. In practice, Runway competes less on raw benchmark scores and more as the tool where generated footage gets edited and finished.

Audio is now standard

The 2026 releases show that synchronized sound is expected, not a bonus. One buyer's guide argues that native audio now separates tools more than visual quality does16. Kling 3.0 already produced dialogue, sound effects, and ambient sound in a single pass, with speech in five languages11. Seedance 2.5 generates 30-second clips with audio and accepts up to 50 multimodal references15. Even the small open-source models now include synchronized audio by default11.

This ties video generation closely to the voice synthesis industry, which is growing quickly in its own right. Twilio's voice AI revenue reportedly grew 60% in its fourth quarter. Telnyx says it can host LiveKit voice agents at half the cost with responses under 200 milliseconds25. Synthesia's Express-2 engine pairs voice cloning with a diffusion transformer to produce full-body avatars that gesture. The company has also launched a plugin that builds videos inside ChatGPT.

The same capability is also enabling fraud. Reporting on voice phishing, known as "vishing," says attackers can clone an executive's voice from short public clips. The result removes the usual warning signs of a scam call, such as an odd accent or unnatural rhythm. Experts recommend out-of-band verification and approval rules that apply regardless of who appears to be calling21. The industry is selling voice realism, and that same realism makes the attacks work.

Image generation keeps accelerating

Still-image generation had its own major release this week. OpenAI shipped ChatGPT Images 2.6 on November 4. It adds animated GIF creation and saved reference images, and it comes with two API models: Flare, built for speed, and Sunburst, built for quality23. OpenAI says generation is about 50% faster. It also says users now create more than 4 billion images a week, up from 3 billion in September23. Both API models cost the same per token, and the API also adds SVG output and markup-based editing28. One platform reported that Sunburst took first place on DesignArena's image leaderboard on release day27.

Other companies are cutting prices. On November 2, xAI retired its older premium image endpoints. Requests to them are now served by a lower-quality setting of Grok Imagine Image 2.0 at a lower per-image price22. One aggregator's chief executive said major image models now arrive every few weeks27.

OpenAI's moves form a clear pattern. It gave up on standalone video but is investing heavily in images, and GIF support takes it a small step back toward motion. This looks like a deliberate choice to compete where it already has scale rather than against Kling and Wan on cinematic video.

Where the coverage disagrees

Much of the reporting comes from fast-moving buyer's guides, and their details often conflict. Sources give Kling 3.0's launch as February 411 or February 54. The early Kling 4.0 Flash preview was described as offering 4K and 30-second clips11 in one guide and 720p and 20-second clips in another13. Pricing comparisons vary a lot depending on route, resolution, and whether audio is included9. At least one recent roundup still recommends Sora 2 to founders, even though the API has been shut down26. Readers should treat any single specification table cautiously.

The takeaway

The big trends are consistent across the coverage, even where the details are not. Clips are getting longer, from about 8 seconds to 30 seconds and, according to some vendors, to several minutes. Audio is now built in. Control is moving from writing prompts to directing with keyframes and references. Distribution is concentrating in aggregators such as Runway, Higgsfield, and HeyGen, which offer models from many developers in one place110.

The next phase looks set to be a contest over workflow and price more than benchmark scores. Agent-driven tools such as Kling's MCP integration14 and Tencent's HyCreator, which is designed to make films about 10 minutes long17, suggest the next competition will be over who controls the production pipeline, not just whose model produces the best clip.