Kling AI 4.0 released: 30 seconds per generation, stereo audio and multilingual lip sync; Flash capped at 720p/20s
The Decoder reported on October 19, 2026 that Kling AI released 4.0: up to 30 seconds per generation, up to ten keyframes, 8,000-token prompts, new stereo audio and lip sync in more than a dozen languages, plus a Flash version capped at 720p and 20 seconds; API pricing and support status are still listed as coming soon.
What changed
According to a report by The Decoder on October 19, 2026, Kling AI released version 4.0. The report says a single generation can produce video up to 30 seconds long, the system accepts up to ten keyframe images to lock in scene transitions, character states, and narrative pacing, and prompt understanding expands to 8,000 tokens. The upgraded Omni Reference accepts up to 15 reference assets at once, including up to 10 images, 5 videos totaling no more than 30 seconds, and 7 elements; storyboards, wireframes, image grids, white-model videos, and depth-map videos can all serve as references. Creative Recreation allows an existing video to be used as a template to replace a product or character, and the report says expressions, body movements, camera angles, or backgrounds can be adjusted individually without affecting the rest of the clip. On audio, it adds two-channel stereo and improves lip sync and dialect accents in more than a dozen languages including German, English, Japanese, Korean, and Spanish; in-video text generation supports multiple languages and stays stable during camera movement. 4K and 1080p 10-bit HDR output and video extension up to two minutes are both listed as coming soon. The simultaneously launched Kling 4.0 Flash focuses on faster generation and lower cost, with a cap of 720p and 20 seconds. The creation interface has been redesigned, and Kling 4.0 can be accessed via MCP and CLI; the report explicitly states that Kling 4.0 API pricing has not been announced and that support status is listed as coming soon. The report also compares Kling 4.0 with Wan 3.0 and Seedance 2.5, saying it holds its own but does not lead, particularly in shot logic continuity, physics, and dialogue naturalness.
What it means for developers
For developers, the most direct change is the expansion of the input and output contract and the preview of access methods. Thirty seconds per generation, 8,000-token prompts, and up to 15 reference assets mean long shots or continuous narratives may no longer require external segment stitching and index alignment, but request bodies are larger and asset preparation and validation logic becomes more complex. Ten keyframes and structured references such as depth maps and white-model videos shift control from pure text prompts to geometry and the timeline, which suits teams that already have storyboards or 3D previsualization assets but requires asset quality and naming conventions to be in place first. Two-channel stereo and multilingual lip sync will affect post-production workflows: if audio is generated together with the video, the traditional pipeline of producing picture first and then dubbing and aligning may need to be re-divided. The Flash version pushes resolution and duration down to 720p and 20 seconds, pointing to batch drafts and preview scenarios. It must be emphasized that MCP and CLI are currently only reported as access methods, and API pricing and support status are both listed as coming soon, so production calls cannot be inferred at this stage; the comparison with Wan 3.0 and Seedance 2.5 in the report is a qualitative media description and does not constitute a reproducible benchmark conclusion.
Integration and evaluation
The following are recommendations and conditional analysis, not reported facts. Recommendation one: if a team already has storyboards or 3D previsualization assets, first compare two paths using the same script, direct generation from pure text prompts versus controlled generation with multiple keyframes, to verify the benefit of keyframe and depth-map controls. The precondition is that reference assets have been cleaned and resolution unified; measurable indicators are the share of shots requiring manual rework and regeneration under the same script, and the pass rate for character and scene consistency between shots. If asset conventions are not mature, continuing to rely mainly on text prompts may be more cost-effective. Recommendation two: if the goal is to embed video generation into an existing automated pipeline, first build a read-only MCP and CLI integration sample based on public documentation, wrapping vendor calls behind your own adapter layer rather than binding the workflow directly to their interface. The precondition is that interface documentation can be obtained reliably; measurable indicators are median time from task submission to obtaining usable assets, and the proportion of failed generations that can be automatically retried successfully without manual intervention. Recommendation three: if multilingual lip sync is a core requirement, build small-scale comparison sets for each target language and treat stereo audio and lip alignment as separate acceptance items. The condition is that the target language falls within the more than a dozen it claims to support; the measurable indicator is the pass rate of native-speaker reviewers on audio-video synchronization, not merely whether generation succeeded. If the target language is at the edge of dialects or accents, it is advisable to experiment first on a low-cost path such as Flash.
Limits and open questions
It should be clear: Kling 4.0 and its Flash version do not currently appear in KimiSeek's catalog of models for sale, making it a version not yet included; even if other versions in the same family are included in the future, that should not be taken to mean this version is available. The API pricing and support status mentioned in the report refer to vendor Kling AI's own plans, whose pricing page shows it has not been announced and lists support status as coming soon. 4K and 1080p 10-bit HDR output and video extension up to two minutes are likewise reported as coming soon and do not mean they are already available. Descriptions such as the 30-second duration, 8,000 tokens, 15 reference assets, and lip sync in more than a dozen languages are all relayed from Kling AI's claims and have not been independently verified; the comparison with Wan 3.0 and Seedance 2.5 is only a qualitative observation in that report, not a standard benchmark. The capability boundaries, rate limits, concurrency, and billing methods of MCP and CLI are not explained in the report, and vendor documentation must be the basis before actual integration. In addition, the judgment about slowing progress in the report is commentary and speculation, and the source itself acknowledges that it must wait for the next generation of Wan and Seedance to be released before judging.