The specification sheets for AI video models have stopped being about whether they can make a shot and started being about whether you can direct one. Kling 4.0, announced on 19 October 2026, is almost entirely a release about control.
What is in it
Generation and input. Up to 30 seconds in a single pass — the number that matters most, because stitching is where continuity dies. Up to 10 keyframe images to pin scene transitions, character states and story beats. Prompt comprehension extended to 8,000 tokens.
Omni Reference accepts up to 15 reference assets per generation, and the accepted types are the giveaway as to who this is for: storyboards, wireframes, image grids, white-model videos and depth-map videos. Those are not consumer inputs. That is a list assembled by asking people who already work in a pipeline what they would hand a model. The per-category limits as reported — up to 10 images, up to 5 videos capped at 30 seconds combined, and 7 “elements” — add to 22 rather than 15, so either the 15 is a per-request cap across categories or the breakdown is imprecise; we could not resolve it against the company’s own notes.
Creative Recreation uses an existing video as a template: swap the product or the character, adjust expressions, body movement, camera angle or background. That is the advertising-production feature, named plainly.
Audio. Two-channel stereo, and improved lip sync across more than ten languages including German, English, Japanese, Korean and Spanish, plus dialects and accents. There is also in-video text generation in multiple languages that is claimed to stay stable under camera movement, which has been one of the genre’s most reliable tells.
Access. A redesigned creation interface, plus MCP and CLI access for agent-driven workflows.
Kling 4.0 Flash is the cheaper, faster variant, capped at 720p and 20 seconds.
Listed as coming soon: 10-bit HDR at 4K and 1080p, video extension to two minutes, and API support. No pricing or credit figures were announced for the 4.0 API.
Stereo is the sleeper feature
Mono audio on generated video is a watermark. Not because anyone consciously hears the channel count, but because every piece of real footage you have ever watched has a stereo field, and its absence reads as recording rather than scene. A model that places sound in a horizontal field is making a claim about the space the camera is in, and that claim has to agree with the picture — which is a harder generation problem than it sounds and a bigger perceptual payoff than another resolution step.
The result nobody wanted to lead with
The Decoder’s Matthias Bastian ran 4.0 against Alibaba’s Wan 3.0 and ByteDance’s Seedance 2.5, and his reading is that Kling 4.0 holds its own without pulling ahead — specifically weaker in shot sequencing, physics and natural-sounding dialogue. His suggestion is that this could be an early sign of AI video progress levelling off, with upcoming Wan and Seedance releases as the test.
That is worth more attention than the feature list. Two years of releases in this category were step changes you could see in a single clip. A release that adds stereo, keyframes, a longer window and an agent interface, and comes out even on quality, is the profile of a maturing product category rather than an advancing research frontier. Which is good news for anyone who has to actually deliver a job with one of these, and a flat note for anyone waiting for the model that solves physics.
Sourcing, plainly
Our reporting here follows The Decoder’s write-up of 19 October. We tried to verify the specifications against Kling’s own release notes, and the page at kling.ai/release-note is client-rendered — it serves a title and the heading “Release Notes” to a fetcher and nothing else, so we have not read the company’s own wording and are not claiming to.
One thing we could verify independently: the official Meet KLING 4.0 video embedded above was uploaded on 28 September 2026, three weeks before the announcement. That matches third-party reports of Flash having been in limited early access from late September, and means the 19 October date is a general-availability date rather than a first appearance.
Sources
- Kling AI drops version 4.0 with stereo audio, multilingual lip sync, and a lighter Flash variant — The Decoder
- Kling AI release notes
- Kling API pricing
- Meet KLING 4.0 | You Call the Shots — Kling AI on YouTube