Kling 4.0 AI Video Generator, Claude Sonnet 5.5 + Opus 5.5, New German AI Robot

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Kling 4.0 AI Video Generator, Claude Sonnet 5.5 + Opus 5.5, New German AI Robot · Channel: AI News · 07:46 · Watch on YouTube

Topics: Claude, Model inference and optimization, Multimodal models, Robotics and autonomous systems

Summary

The video covers Kling 4.0’s nine new video-generation features, including longer clips, multi-key-frame control, expanded audio and language support, and video extension. It then reviews Claude Sonnet 5.5, describing its benchmark results, speed and token efficiency, pricing, and early tester reports. The video also discusses ways users can reduce Opus 5.5 costs by routing messages, before introducing Munich-based humanoid robotics company Vesoma and its self-trained walking prototype. It closes with a mention of an unboxing of Limx’s Luna humanoid.

Key points

Kling 4.0 supports up to 10 key-frame images and single generations lasting 3 to 30 seconds.
Its prompt limit has increased to 8,000 tokens, and the model adds two-channel stereo audio with improved lip sync.
Kling 4.0 adds 10-bit HDR, native 21:9 ultrawide support, German, French, and Indonesian, plus legible on-screen text, emojis, and logos.
Video extension can stretch a generation to two minutes total, but the feature is marked as coming soon.
Kling retains text-to-video, image-to-video, first-to-last-frame control, Omni reference, and video editing features.
Anthropic presents Claude Sonnet 5.5 as a faster, cheaper complement to Opus 5.5; the video cites competitive benchmark results and efficiency gains.
Sonnet 5.5 pricing remains $2 per million input tokens and $10 per million output tokens, according to the video.
The video says users can lower Opus 5.5 costs by routing messages through a tool that filters requests before calling the model.
Munich-based Vesoma says it developed a walking humanoid prototype in six months using a self-taught policy and aims to serve logistics, manufacturing, warehousing, and care work.

Chapters

Start	Chapter	Summary
00:00	Kling 4.0’s new video features	The video details Kling 4.0’s nine additions, from key-frame control and longer generations to new formats, languages, and video extension.
03:49	Claude Sonnet 5.5	The video compares Sonnet 5.5 with Opus 5.5 and covers benchmarks, efficiency claims, pricing, and early tester reports.
06:01	Lowering Opus costs and a new German humanoid	The video discusses message routing to reduce Opus usage, introduces Vesoma’s humanoid prototype, and mentions an unboxing of Limx’s Luna.

Quotes

Kuaishou didn’t just update Kling, they rebuilt it with nine new features. — 00:00

prompt length, which jumped to 8,000 tokens. — 01:05

allowing creators to stretch a generated video across multiple extensions up to 2 minutes total. — 03:15

the model runs more than 30% faster than Sonnet 5 while using roughly a third fewer tokens for the same tasks — 04:53

learning physical work through trial and error instead of imitating human demonstrations. — 07:06

Concepts and entities

Kling 4.0, Kuaishou, Claude Sonnet 5.5, Claude Opus 5.5, Anthropic, Terminal-Bench 4.0, GDPval, OSWorld 2.1, Jev, Vesoma, Martin Reed Miller, LimX Luna

Full transcript (original language)

00:00 Today on AI news, Kling 4.0 drops nine new AI video generation features. Claude Sonnet 5.5 shakes the industry and state of the end for a brand new German humanoid. But first, Kuaishou didn’t just update Kling, they rebuilt it with nine new features. Starting with one of the biggest, number one, multi key frame generation. Up to 10 key frame images now guide a single generation, giving far more precise control over how a shot’s story beats land, rather than leaving the model to guess what happens between one moment and the next, which

00:33 matters most in longer multi-seene shots where continuity used to break down. And that control structure only matters if the shot has room to actually play out. Which brings us to feature number two, video duration. as now a single generation can run between 3 to 30 seconds instead of Clling’s old short clip ceiling. So, a full scene can unfold in one continuous take instead of getting stitched together from several shorter generations. But longer generations need more directions to fill it. And that’s where feature number

01:05 three comes in, prompt length, which jumped to 8,000 tokens. So, more detailed multi-part direction covering character actions, camera moves, pacing, and dialogue all in one prompt. actually gets read and followed instead of getting cut off partway through. And once the model can take in the full prompt, the next upgrade is about to convincingly execute it. With feature number four, which is two channel stereo audio, replacing the single flat audio track from before, paired with what Quaisho calls far more accurate lip

01:35 sync. In fact, Kuaishou frames this as its biggest leap yet in combined audio and visual quality with faces holding detail through motion, lighting behaving more realistically, and body language carrying real emotion instead of stiff generic gestures. And that visual leap carries straight into color itself with feature number five 10-bit HDR arriving at both 1080p and 4K marked as coming soon for richer color and smoother transitions between light and shadow, especially in high contrast scenes like backlit shots, night driving or neon lit

02:09 streets with noticeably less color banding along the way. And alongside richer color, Kling is also rethinking the frame itself with feature number six, native 21x9 ultrawide support, so that cinematic theater style aspect ratios don’t need to be faked or cropped in post anymore. And this is on top of the already supported 16x9, 1x1, and 9x6 formats. Additionally, the following two features are about who Kling can actually speak to, with feature number seven adding three new languages, German, French, and Indonesian. And

02:42 because of this, Kling now supports nine total languages. But feature number eight also deals with language as Kling can now generate legible on-screen emojis and logos as part of its text generation, not just words. With Kuaishou saying that text stays readable even as the camera moves or the scene cuts. And rounding out the list is the one change that circles back to runtime just from the other direction, which is feature number nine, video extension, allowing creators to stretch a generated video across multiple

03:15 extensions up to 2 minutes total. But something to note is that this feature is marked as coming soon. And in addition to this, what Kling 4.0 brings to the table in terms of new features. It’s also packing the same powerful fundamentals, including text to video, image to video, and first to last frame control, all carried over as is with resolution topping out the same as Kling 3 at 720p, 1080p, and 4K. And it also features the same Omni reference system which lets one generation draw up to 15 total reference assets, up to 10 images,

03:49 up to five videos totaling 30 seconds, and up to seven subjects with a specific voice assigned to each one. And video editing supports up to five input clips per edit. But while Kling 4.0 0 is pushing the envelope of AI video generators. Anthropic is pushing in a different direction with the release of Claude Sonnet 5.5, framing it as a faster, cheaper complement to Claude Opus 5.5 for everyday coding, bug fixes, and document work rather than a flagship replacement. And on Terminal Bench 4.0, know a benchmark for agentic coding in a

04:21 real terminal. Sonnet 5.5 actually edges out Opus 5.5 and represents a massive jump over Sonnet 5’s own prior result. And on GDP vala measuring performance on realworld professional tasks, Sonnet 5.5 comes in just barely behind Opus 5.5, essentially matching its own larger sibling. And on OSWorld 2.1, a benchmark for operating a computer, Sonnet 5.5 lands just under Opus 5.5’s result. Anthropic also says it’s the first Sonnet model to beat Pokémon Red using

04:53 only screenshots, a longrunning internal test of multi-step reasoning over Long Horizons. Plus, Anthropic says the model runs more than 30% faster than Sonnet 5 while using roughly a third fewer tokens for the same tasks, cutting per task cost by up to 30%. And at a low or medium reasoning effort, Anthropic says Sonnet 5.5 can match Sonnet 5’s best ever performance for roughly onetenth of the cost per task. But even more importantly, pricing itself hasn’t moved. It’s still $2 per million input tokens, $10 per million output tokens,

05:27 which is the same rate as Sonnet 5, and half of the cost of Opus 5.5. And early testers are reporting impressive stats with Box reporting 2.4 four times faster runs with 12% fewer tokens. Zenesk reporting 20% faster ticket processing and Slack reporting fewer output tokens and fewer steps to reach the same result. In fact, coding platform Lovable reported roughly a third fewer tool calls and about half as many shell executions per task. Anthropic also says the model matches Opus 5.5 security safeguards and adds classifiers meant to

06:01 block attempts to extract its reasoning process. But this efficiency push isn’t entirely limited to sonnet 5.5 as users are also cutting their opus 5.5 bills by over 35% by simply routing messages through type safes jev. This is because Jeb sorts each message by type urgency and whether it needs a real reply for a fraction of a scent so that opus 5.5 only gets called in when the job actually needs hard reasoning. But embodied AI is also leaping ahead as another humanoid robot company from Munich, Germany is making strides and

06:36 its name is Vesoma, founded in December 2025 and it just came out of stealth. In fact, the company has grown to over 60 employees, filling more than 3,500 m of lab and office space with former Google DeepMind researcher director Martin Reed Miller now leading AI. The Soma says it went from nothing to a walking humanoid prototype, moving on a policy that it taught itself in just six months, and it calls its approach physical general intelligence, learning physical work

07:06 through trial and error instead of imitating human demonstrations. Voma says it’s building its humanoids to work in logistics, manufacturing, warehousing, and eventually care work. But make sure to tell us down below in the comments how much you think this robot should cost. And finally, check out this unboxing of Limx’s brand new Luna humanoid.

07:46 Anyways, like and subscribe for more of the latest in AI and robotics news and thanks for watching.