Three things happened in AI that are worth a few minutes. Kling pushed its video clips to 30 seconds. A collaboration between fal and Black Forest Labs produced a video model that fal says generates 10 seconds of footage in 12.45 seconds. And ElevenLabs shipped a new model family for voice agents that handles the LLM, speech-to-text and text-to-speech settings so developers don’t have to wire them up themselves.
Kling 4.0 Is Out, and It Generates 30-Second Clips
Kuaishou’s AI video tool, Kling, launched Kling 4.0 on October 15 with a headline change that short-form video creators will notice immediately: native generation up to 30 seconds in a single run (Source). Kling says the new version delivers “more stable, physically convincing action in demanding scenes, from fast-paced fights to sports and large-scale movement.” The company also says stronger camera direction “supports dynamic tracking, complex camera moves, and more coherent shot composition.”
The new model introduces Omni Reference, which supports up to 15 combined reference assets, including images, videos, and subjects. Kling says creators can combine character, scene, motion, and style references in the same generation and try different combinations “without rebuilding your entire creative setup.” On top of that, Kling 4.0 supports up to 10 flexible keyframes, which Kling says give “additional control over composition, character actions, and scene progression.”
Kling’s homepage bills the update as “Audio-Visual Performance for Realistic Results,” promising “an immersive audio-visual experience, taking video generation to professional-grade production.”
The practical question for any creator right now is whether 30 seconds of coherent AI video opens new content formats, or whether it just delays the edit. My bet: for a lot of short-form pieces, one generation now covers a meaningful chunk of the finished cut.
Why it matters: For content creators and social teams, 30 seconds is not a minor spec bump. Shorter clips mean generating several pieces and stitching them in post, which adds time and introduces seams. If a single 30-second generation holds its motion and camera work across the whole duration, that changes the economics of AI-assisted short-form video. It doesn’t make the tool free, and the output still needs review, but the assembly step gets shorter.
Find out more: Kling 4.0 on kling.ai
FLUX 3 Max: A Video Model Built for Speed Without Sacrificing Quality
This month, fal Research launched FLUX 3 Max, a post-trained video generation model built in collaboration with Black Forest Labs on its FLUX 3 foundation. The pitch is speed at quality: fal says FLUX 3 Max generates a 10-second video in 12.45 seconds. For comparison, Seedance 2.5 takes 277.80 seconds for the same output. fal calls that 26.5 times faster, with 96.2% lower latency, in its own benchmark measurements.
On quality, fal says FLUX 3 Max sits “statistically tied with the highest-ranked state-of-the-art video models” in image-to-video evaluation. Black Forest Labs’ own homepage describes the model as “22× faster than Seedance 2.5 at the same quality.” That figure is what you get by dividing fal’s own two timings (277.80 ÷ 12.45 ≈ 22.3); fal’s blog does not explain how it arrives at 26.5×.
The model is available now on fal for text-to-video, image-to-video, and keyframes-to-video generation. The FLUX 3 family, as Black Forest Labs describes it, also supports optional audio alongside video generation, including multilingual speech, effects, and ambience. fal’s Max announcement focuses on speed and quality in video generation specifically.
This is a collaboration between a foundation model lab (Black Forest Labs) and fal Research, with fal handling the post-training and inference optimization on top of BFL’s FLUX 3 foundation model. fal says it applied the same approach earlier with H3 Max, which it post-trained on MiniMax H3 open weights.
Why it matters: Speed in video generation is not just a convenience feature. At 22× to 26.5× faster than Seedance 2.5, depending on whose figure you use, FLUX 3 Max moves from a tool you queue and come back to a tool you can iterate with. For a content team or a solo creator who needs to test three concepts before noon, the difference between 12 seconds and 278 seconds is the difference between done and waiting. This model makes fast iteration on AI video practical, which is different from just making it technically possible.
Find out more: Introducing FLUX 3 Max by fal
ElevenLabs Ships Eleven Realtime: Voice Agents That Reason for You
On October 19, ElevenLabs added Eleven Realtime to its changelog. It is a new family of models for ElevenAgents voice agents, and the core idea is that ElevenLabs handles the configuration you used to wire together yourself. From the changelog: “You pick one model, and ElevenLabs handles the LLM, Speech to Text and Text to Speech settings behind it. Your agent’s system prompt, voice, tools and knowledge base work the same as before.”
Two models are available. Eleven Realtime v1 Mini is the default: the changelog calls it “the fastest and cheapest,” and ElevenLabs’ product page aims it at AI assistants, educational tutors, and interactive characters. Eleven Realtime v1 Max is the reasoning variant. The changelog describes it plainly: “Eleven Realtime v1 Max reasons before it answers, so use it when your agent has complex instructions or multi-step tool calls.” ElevenLabs’ product page lists time to first audio at 570ms for Max and 620ms for Mini.
ElevenLabs says Max scored 98.8% on Big Bench Audio, a benchmark that evaluates reasoning through spoken questions. ElevenLabs does not list a Big Bench Audio score for Mini.
For developers already building on OpenAI’s Realtime API, the switch is designed to be low-friction. The changelog notes the events and WebSocket transport are the same; changing the endpoint and session configuration is the migration path.
The models are rolling out gradually across ElevenAgents workspaces. For earlier context, see our ElevenLabs voice model coverage.
Why it matters: Building a working voice agent is not just a TTS problem. The hard parts are the decisions the agent makes mid-conversation: when to respond after a pause, what to do when the user interrupts, when a “mm-hm” means the user is still listening versus finished. Developers have been solving those coordination problems themselves, or piecing together component models. Eleven Realtime packages that configuration. Whether it actually performs better in production than a carefully tuned custom stack is a separate question. But for a team that wants to ship a voice agent without becoming experts in turn-taking model selection, the offer is worth evaluating.
Find out more: ElevenLabs changelog, October 19
For more on the AI video creator space, see our roundup of what practitioners are saying about AI video tools in 2026. And if you’re evaluating the FLUX family for image generation, FLUX AI pricing in 2026 covers the full cost breakdown.