AI video had a dramatic September. OpenAI shut down the Sora 2 API on the 24th — calls now return 410 Gone, with no successor named. Google made its Vids editor free in 1080p on the 23rd. And on September 30, Kling 4.0 entered early access, pushing the length frontier that everyone in this market is chasing: native 30-second generation in a single pass.
Kling, from Chinese company Kuaishou, has been the quiet workhorse of AI video — strong on human motion, fast, and affordable. Version 4.0 is its bid to lead the next phase. Here is what is new, what it costs in practice, and how it stacks up.
Thirty seconds changes the grammar
The headline spec: Kling 4.0 generates up to 30 seconds of video natively, double the 15-second ceiling of Kling 3.0. That sounds like an incremental number. It is not.
Video length determines what you can say in a single shot. Fifteen seconds is a reaction, a product spin, an establishing beat. Thirty seconds is a scene — enough for a setup and a payoff, a before and an after, a full joke. Every extra second of coherent generation also compounds the difficulty: characters must stay consistent, motion must stay plausible, and the model has to hold its plan together across roughly twice as many frames. Native 30-second generation, rather than stitched-together 15-second segments, means fewer visible seams and fewer continuity breaks.
Kling is not alone at 30 seconds. ByteDance’s Seedance 2.5 also runs to 30 seconds (at 720p), and the open-weights Wan 3.0 reaches 30 seconds too. But Kling 4.0 pairs the length with a control package aimed at professional workflows:
- Up to 10 keyframe images — direct the shot’s major beats instead of hoping the model invents them.
- Up to 15 reference assets — keep characters, products, and styles consistent across the clip.
- Two-channel stereo audio — sound designed in the generation, not bolted on afterward.
The Flash tier and the rollout
Kling 4.0 Flash opened to Ultra Yearly subscribers on September 28, two days before the standard 4.0 early access began on the 30th. The wider rollout is described as “coming in October” — which, as of this writing, means now. Early access with a staggered rollout is the standard pattern: reward the highest-paying tier first, gather performance data, then open the gates.
The state of the video market it enters
Kling 4.0 arrives into a market that just lost its most famous name. The Sora story is worth recapping because it explains the opening Kling is walking through:
- April 2026: OpenAI discontinued the Sora web and iOS app. The model lived on inside ChatGPT subscriptions and the developer API.
- September 24, 2026: OpenAI removed
sora-2andsora-2-profrom its Videos API. Calls return 410 Gone. OpenAI’s own documentation lists no one-to-one substitute. - September 23, 2026: Google made Vids free — 1080p AI video generation and upscaling for anyone with a Google or Workspace account, running on Gemini Omni 1.1 Flash, every clip carrying an invisible SynthID watermark.
That leaves the field to Google’s Veo 3.1 (cinematic realism, native audio, still the reference for prompt adherence), Runway’s Gen-4.5 (creative control — Motion Brush, Camera Control — and now bundled with Kling 3.0 Pro access), Seedance 2.5 (stylized motion, 30 seconds at 720p from 34 credits on aggregator platforms), MiniMax’s H3 Max (open weights, 2K video with native synced stereo audio, renders faster than realtime), and the open-weights Wan 3.0 (30 seconds, cheapest per-clip costs in the business).
Kling’s historical edge has been human motion — bodies that move believably, which remains the hardest thing in video generation — plus speed and price. Version 4.0 keeps that positioning and adds the length and control that professional users kept asking for. The 10-keyframe support is the sleeper feature: it moves the tool from “prompt and pray” toward actual direction.
What “early access” really means for you
Early access in AI video usually means three things: limited capacity, evolving quality, and pricing that will change. If you are evaluating Kling 4.0 for production work, test the failure modes, not the demos. Long generations fail in specific ways — faces drifting over 30 seconds, hands doing impossible things in the back half, audio desyncing from motion — and those failures only show up in your prompts, not the launch reel.
The reference-asset system deserves particular testing. Character consistency across a 30-second clip is the feature that would make Kling 4.0 genuinely useful for campaigns and series work — the same job that aggregator platforms attack with systems like Soul ID, which locks a face across generations. If Kling’s 15 reference assets deliver that natively, it collapses a whole post-production workflow into the prompt.
The audio question
Two-channel stereo in Kling 4.0, native synced stereo in MiniMax H3, native audio in Veo 3.1 — sound is no longer an afterthought in video generation. It is becoming a first-class output, which makes sense: half of what makes video feel real is audio. For creators, this simplifies pipelines enormously. For the voice-AI world, it is convergence — the same week voice models got dramatically better, video models started generating their own soundtracks. Expect the boundary between “voice AI” and “video AI” to keep dissolving.
The bottom line
Kling 4.0’s pitch is simple: the longest single-shot generations in its class, with director-level controls, from the lab with the best reputation for human motion. The 30-second ceiling, 10 keyframes, 15 reference assets, and stereo audio add up to a tool aimed at people making actual content, not demos.
The open question is the same one facing every video model in October 2026: reliability at length. Thirty seconds of coherent, controllable video — generated fast, priced sanely — would be a genuine step change for creators. Whether Kling 4.0 delivers it consistently, or only in the launch highlights, is what the next few weeks of real usage will reveal. For the image-generation side of the creative stack, see GPT Image 2.5’s two-tier launch, and for the lab trying to unify video, audio, images, and robotics in a single model, read about FLUX 3.
Featured image: Jud McCranie via Wikimedia Commons (CC BY-SA 2.0)
