AI video generation: Kling 4.0 ships 30-second clips globally

The latest move in a crowded field

Kuaishou's Kling AI has released Kling 4.0, its newest video generation model, worldwide. The company is pitching it on three improvements: more realistic output, finer creative control, and better narrative coherence.11 Kuaishou's Hong Kong filing says the model officially launched on October 15. The English-language releases came out a day later.1213 The main upgrades are:

• Stabler motion: fast action and complex camera moves hold together better.
• Longer clips: a single generation can now run up to 30 seconds.
• Omni Reference: creators can combine up to 15 reference assets with up to 10 keyframes.1112

The release matters less as a single product and more as a sign of where text-to-video has been heading all year. Longer single takes, heavy use of reference material, native audio and output built for professional pipelines have become the standard feature list. Kling 4.0 is the latest model to reach that standard, not the first.

What Kling 4.0 actually changes

Kuaishou is aiming at what it calls a long-standing weakness of AI video: images that fall apart during high-speed motion or sweeping tracking and orbit shots.12 Seoul Economic Daily reported the same focus. It said the company claims continuity holds even when the camera follows a person or circles a subject.13 Kling also has an extension feature that can stretch an uploaded video to two minutes, which supports stories told in several parts.11

The launch also puts weight on agents and developers. Kling 4.0 connects to Kling's MCP and CLI tools, so AI agents can call its reference and keyframe-to-video features. Kuaishou says enterprise and developer API access is coming soon.12 The company says it has more than 100 million users and over 50,000 enterprise customers in advertising, film, TV and gaming.11 A separate industry roundup gives a different figure: more than 60 million creators and 600 million videos since Kling's mid-2024 debut.15 The two numbers probably measure different things, so neither is clearly wrong.

For marketing, Kuaishou partnered with director Kim Sung-su, whose AI short film made with the model was shown at the Busan Film Festival.13

Where the coverage diverges

The reports don't agree on which features are available now and which are still coming. Kling's global press release says 10-bit HDR output at 4K and 1080p, plus a 21:9 ultrawide format, is coming "soon."11 Seoul Economic Daily says 10-bit HDR is supported now and that 4K and 21:9 are planned.13 An investment-news summary says HDR and 4K are both supported.14 When Kling first announced Kling 4.0 on September 28, Forbes reported the 4K, HDR and 21:9 output as part of the model.8

The company's own wording is the most cautious version, so buyers should treat the 4K HDR output as promised rather than proven until it actually ships. Clip length and reference limits are the features to rely on today.

The launch was also gradual. A smaller Kling 4.0 Flash went to a limited group of early-access users on September 28, and the full model followed in mid-October.15 That pattern of preview first, then wider release, has become common in this market.

A spec sheet everyone now shares

Read side by side, Kling's headline features look a lot like everyone else's.

• ByteDance Seedance 2.5 launched July 31 with 30-second audio-video clips and multiple rounds of extension.16
• Alibaba's Wan 3.0 also produces 30-second clips and accepts 20 references.15
• Vidu Q4 Preview from ShengShu Technology came out October 7 with up to 15 image references, three audio references and 10-bit color at 2K or 4K.17
• Luma's Ray3.2 already accepted up to 16 keyframes and produced native 16-bit HDR.15

Forbes made a similar point: competing labs are working on the same problems, and MiniMax's H3 takes text, image, video and audio as input.8

The result is that being first with a feature no longer sets a model apart. Price and access matter more. Vidu's preview starts at $0.014 per second.17 HeyGen's new general-purpose video model, built on MiniMax H3, was priced at $0.01 per second on its API through October.20 Open-weight options keep growing. Kandinsky Lab released Kandinsky 6.0 Video under the MIT license on October 6. It comes in two sizes and generates five-second clips with synchronized 44 kHz audio.7 Developers are already shrinking MiniMax H3 so it runs on 8 GB consumer GPUs.7

Leaderboards that don't agree

The industry has no agreed answer on which model is best. On the Arena text-to-video ranking, Google's Gemini Omni 1.1 Flash led as of October 14, statistically tied with Gemini Omni Flash and ahead of Black Forest Labs' FLUX 3 Video.2 Artificial Analysis puts Alibaba's Wan 3.0 first on text-to-video, both with and without audio.1 One review site recently made MiniMax H3 its overall video pick because it scores near the top on several boards without winning any single one.6 A smaller blind-vote leaderboard with about 1,450 votes has Kling v3 in first place.5 Another review says Kling 3.0 doesn't even reach the top 16 on Arena's text-to-video board.6

These gaps come from different methods, voter pools and test prompts, and rankings move roughly every week.1 The practical lesson for buyers is that a model's leaderboard rank says less than its fit for a specific job, such as long takes, image-to-video or self-hosting.

The Sora shutdown

The biggest strategic story of 2026 in AI video has been OpenAI stepping back. OpenAI shut down the Sora web and app experience on April 26 and lists September 24 as the date the Sora API was discontinued.16 One marketing-focused overview noted the app closed about seven months after launch.20 OpenAI may not be gone for good. The Information reported that it is discussing giving a filmmaker-focused startup access to Sora so the startup can sell software and services built on it.9

Google has moved in the other direction and folded video into a general model. It introduced Gemini Omni in May as a model that can create from any input, starting with video. Omni Flash is now the default video model in Google's API documentation, while Veo 3.1 handles specialist tasks like scene extension.15 Google says Omni will replace Veo in the Gemini app.20 The split between OpenAI dropping a standalone video product and Google building video into its general multimodal model is a useful way to read the market. In this reading, standalone video models increasingly belong to specialists, mostly Chinese platforms, while large labs treat video as one output among many.

Voice synthesis is following the same path

The same push toward "directable" output is happening in AI voice. On October 13, Fish Audio released Drama 3, a text-to-speech model built to be directed in plain language. It can shift emotion partway through a sentence and produce scenes with several characters in one pass.28 Its technical report describes a 35-billion-parameter mixture-of-experts model that activates 3 billion parameters per token and can edit a single word inside existing speech. The weights are public.25 Startup Fortune reported that Drama 3 is free for companies with under $500,000 in annual revenue. It also reported that Artificial Analysis ranks Drama 3 ahead of ElevenLabs' Eleven v4 on pronunciation accuracy.

The large platforms are releasing voice models quickly too:

• Google released Gemini 3.8 Flash TTS and Flash-Lite TTS on September 23. It called this its third audio-model launch in under a month and highlighted designing voices from text prompts in more than 100 languages.23
• Microsoft AI released MAI-Voice-2.1 and a low-latency Flash version on October 1.4
• xAI released Grok Voice 3.0 on October 14.3
• OpenAI began rolling out GPT-Live-1.1 voice models in ChatGPT on October 15.4

Speech tools and video tools are starting to look alike. Both now offer reference casting, natural-language direction, multi-speaker scenes and editing in place. Kling 4.0 itself supports more languages, accents and dialects for spoken output.11

Multimodal models and images

Underneath these releases, general-purpose models keep taking in more kinds of input. On October 15, Meta released open weights for Muse Spark. It is an 816-billion-parameter mixture-of-experts model that accepts text, image, video and audio.4 In images, one tracker reports that ChatGPT Images 2.5 holds the top two spots on every board it follows.10 Google shipped Nano Banana 2.1 on October 6.6 On the open side, an 8-step Qwen-Image-2.1-Turbo appeared in the same week's community roundup.7

Our read

Kling 4.0 is a solid release, but its importance is that it confirms where the market is going more than any single feature it adds. Thirty-second takes, many references, keyframe control and HDR are now basic requirements, and Kuaishou had to deliver them to keep its large user base. Competition now centers on price per second, API and agent access, and whether output can go straight into professional editing. Video, voice and image generation are converging on the same model of AI as a tool the creator directs, and much of the fastest progress is coming from Chinese labs and open-weight releases rather than the US companies that started the category.