Kling 4.0 from Kuaishou: clips up to 30 seconds, 15 references and stereo sound

⚡ The gist in 5 seconds

• Kuaishou has released Kling 4.0: a single pass produces a clip from 3 to 30 seconds, while version 3.0 capped at 15.
• Inside: up to 10 keyframes anywhere on the timeline, 15 references (images, videos, characters, voice), stereo sound with lip sync, prompts up to 8,000 tokens, and launching from AI agents via MCP and CLI.
• The Kling website opens from Russia, registration is free. Russian bank cards don't go through, while the Basic plan adds a watermark and prohibits commercial use.

🔍 What happened

Kuaishou rolled out Kling 4.0 — according to the write-up on Habr, this happened on October 15, 2026. The lighter 4.0 Flash arrived earlier, on September 28, and was available only to holders of the annual Ultra subscription. Extension to 2 minutes is promised later. The resolution wording is careful: Kuaishou writes that the model "will also support" 4K and 21:9. Flash is capped at 720p, while 10-bit HDR in 1080p and 4K will come later.

Photo animation looks like this. A front-facing shot, face close-up, no beauty filters or aggressive upscaling. The photo is loaded as the first frame of the clip. The face is added as a separate character reference. Then a short prompt about calm natural movement. You generate 5 seconds, check the face, and extend only after verification. The first successful take takes 10–15 minutes — if you don't chase 30 seconds right away. The pricing benchmark carried over from 3.0: a second in 720p without sound costs 6 credits, in 1080p with sound — 12. For the new version, the developer hasn't officially published the cost and advises checking it in the interface before launching.

🧠 Our take

The jump from 15 to 30 seconds plus stereo sound with lip sync is a genre shift. An animated photograph used to be a keepsake GIF. Half a minute with speech is already a scene you can drop into an ad or a social media clip. Our assessment: Kuaishou isn't fighting for pretty demos but for a working tool, and the release numbers are tuned to that.

The most interesting part here is the 15 references and 10 keyframes. The main pain of video generation isn't resolution but the face drifting from frame to frame. The source directly names the cause of the "plastic" face: frames are assembled anew, small errors accumulate, and the longer the clip, the worse it gets. Keyframes are exactly what sets anchor points. We believe it's character consistency, not 4K, that will decide whether a clip is fit for production.

Who wins? Aggregators with payment in rubles and via SBP. The write-up itself was authored by the SpeShu.AI blog — a company that sells just such an aggregator: Kling 4.0 there is bundled in one window with 300+ models, and you pay for the clips you generate. It's convenient, but also a reason to read the text with the conflict of interest in mind. Who loses — those willing to wait for direct access and pay with a card that Russians don't have.

⚠️ What we don't know

The per-second price for 4.0 hasn't been published; all credit benchmarks are from the previous version. 4K and 21:9 are promises, not launch facts. The 10–15 minute generation time is a blog estimate, not measurements. The October 15, 2026 date is given as is; we haven't verified it against Kuaishou's documentation. And most importantly: this is material from a company that sells access to this very model.

Prediction: within a quarter, Kuaishou will reveal the per-second price of Kling 4.0 and either confirm or push back the 4K and 21:9 timelines — without a published price, running the model at scale is impossible.

Disclaimer: release details are taken from the SpeShu.AI company blog; we haven't checked Kuaishou's official documentation.

Source: Habr / SpeShu.AI