Kling 4.0 lets you pin up to 10 keyframes and generate 30 seconds in one go. That moves the creative decision back to where it belongs — the storyboard, approved before anything moves. If your team still briefs AI video with a paragraph of prompt, this is the release that makes that expensive.
Kling 4.0: ten keyframes, 30-second shots, 15 references
On 28 September Kling announced Kling 4.0 for October and opened Kling 4.0 Flash to Ultra Yearly subscribers the same day. The list: up to 10 keyframes, native 30-second generation, up to 15 multimodal references, video extension, stereo audio with more accurate lip sync, up to 4K and 10-bit HDR, and support for multiple languages, accents and dialects (Kling AI, 28 Sep).
My read: On the Christmas greeting film I worked stills-first: the images came from Nano Banana, then I built each sequence in Kling and cut it in CapCut. Ten keyframes turn that habit into the tool’s own grammar: the frames you approve are the film. The 30 seconds matter for a different reason. Tiny runs 41 seconds, and the hardest part was one cut — the live-action kitchen and the animated world had to share lighting direction and colour temperature. Fewer joins, fewer places for that to break. The pack is still the thing to check frame by frame. The limit: Flash is a beta on the top annual plan, the full model is “October”. Test one spot. Do not rebuild a pipeline around a beta.
ElevenLabs v4 speaks 90+ languages in one voice
ElevenLabs released Eleven v4 and v4 Turbo on 28 September. Both support more than 90 languages; a voice recorded in one language can speak the others with a native accent while keeping its identity, and Instant Voice Clones now need about 10 seconds of audio (ElevenLabs, 28 Sep). Arabic and Persian are on the supported list; the Style and Speed sliders are not available in v4 and SSML is not supported (ElevenLabs docs). Every clone requires verified consent from the voice’s owner (ElevenLabs).
My read: One brand voice across markets is now a production choice, not a casting problem. That makes the contract the real deliverable. When you clone a voice, write the languages into the consent — not just the duration. And no SSML means you cannot force a pronunciation. Brand names are exactly what a model mispronounces. Listen to every product name in every language before the grade.
Adobe: AI traffic to US retail sites up 130% this holiday
Adobe’s holiday forecast, published 28 September, expects US online sales of $275.1 billion for 1 November–31 December, up 6.7%. It expects traffic from AI chat and browser services to US retail sites to rise 130% year on year — 159% on Thanksgiving. In its July survey of 5,000 consumers, 77% of people who had shopped with AI felt more confident in the purchase and 69% said they were less likely to return it (Adobe, 28 Sep).
My read: This visitor arrives already persuaded. The chat did the comparing; the product page only has to confirm it. So the pack shot, the claim and the size on the page must match what the assistant just said, word for word. A mismatch is where that 77% confidence turns into a bounce. These are US numbers, not Gulf or Iranian ones — borrow the question, not the figure.
What I’d do today
Take your next AI video brief and replace the prompt paragraph with keyframes: five to ten approved stills, in order, with the pack locked. If a still cannot be approved, the shot is not ready to generate. If you want help setting that up, here is how I work.
Sources
- Kling AI — Meet KLING 4.0 | You Call the Shots (official channel), 28 Sep 2026 — youtube.com
- ElevenLabs — Eleven v4: our most expressive text-to-speech model yet, 28 Sep 2026 — elevenlabs.io
- ElevenLabs documentation — Eleven v4, accessed 29 Sep 2026 — elevenlabs.io
- ElevenLabs — Eleven v4 and v4 Turbo, accessed 29 Sep 2026 — elevenlabs.io
- Adobe — U.S. holiday shopping season to hit record $275.1 billion online, 28 Sep 2026 — news.adobe.com