Kling AI ships video generator 4.0: 30-second clips, 8,000-token prompts

Kling AI has released version 4.0 of its video generator, and the numbers on the spec sheet move the product forward: clips up to 30 seconds, ten keyframe images as input, and prompt comprehension expanded to 8,000 tokens. The release also carries a lighter Flash variant, stereo audio, and multilingual lip sync, according to the company's announcement reported by The Decoder.

The performance picture is more complicated. In comparisons against Wan 3.0 and Seedance 2.5 — two competing AI video generation models — Kling 4.0 "holds its own but doesn't pull ahead," The Decoder reports. That result matters beyond one product cycle. It could be an early sign that the breakneck pace of progress in AI video models is starting to level off, after a period in which each major release delivered visibly large jumps in output quality.

What does version 4.0 actually change?

The concrete upgrades are quantifiable:

• Clip length: support for generated videos up to 30 seconds
• Keyframe input: up to ten keyframe images accepted as input
• Prompt comprehension: expanded to 8,000 tokens, meaning far longer and more detailed text instructions
• Audio: stereo sound generation
• Lip sync: multilingual synchronization between generated speech and on-screen mouths
• Flash variant: a lighter-weight model option alongside the main release

Each of these addresses a specific production constraint. Longer clips reduce the stitching work required to assemble multi-shot sequences. Ten keyframes give creators finer-grained control over how a generated video evolves frame to frame. An 8,000-token prompt window allows instructions dense enough to specify scenes, camera behavior, and timing in a single request rather than across several.

Stereo audio and multilingual lip sync point at a different ambition: making generated clips usable as finished media rather than silent footage that requires post-production. Localization is the explicit target here — one generation pipeline that can plausibly serve audiences in multiple languages.

Why the benchmark result is the real story

The comparison data is where the release stops being a routine feature update. Kling 4.0 competed against Wan 3.0 and Seedance 2.5 and did not establish a lead, per The Decoder's testing summary. It held competitive ground. Nothing more.

That is a data point worth watching. AI video has progressed through a stretch of steep, visible gains — each generation resolving artifacts the previous one could not. A release that matches but does not exceed its rivals fits a pattern familiar from other AI categories, where capability curves steepen before they flatten and differentiation shifts from raw quality toward features, cost, and workflow fit.

One release does not confirm a plateau. The Decoder itself frames the finding cautiously, as something that "could be an early sign" of leveling off. But if the next competitive cycle repeats this pattern, the implication for buyers is that model choice will hinge less on benchmark supremacy and more on the practical parameters — clip length limits, audio support, prompt capacity, and variant pricing such as the new Flash option.

What comes next

Kling 4.0's feature set — 30-second clips, ten keyframes, 8,000-token prompts, stereo audio, multilingual lip sync, and the Flash variant — is now the baseline the company must defend against Wan 3.0 and Seedance 2.5. Whether the next round of comparisons shows a genuine capability plateau in AI video, or Kling 4.0 merely caught a strong field on an off cycle, will be decided by the benchmarks that follow.