Mitsubishi reveals a new AI humanoid robot, "N," powered by Highlanders' Kepler v1.0 — a dual-brain AI model designed to plan and react in real time at once. We break down its architecture, the data flywheel behind it, and Highlanders' new manufacturing deal with Mitsubishi, targeting 1,000 units a month by 2027 to help address Japan's shrinking workforce. Then, Boston Dynamics' Atlas shows off perfectly synchronized robot choreography, and Tsubame Industries' piloted mecha robot, Archax, brings sci-fi to life as a real, working machine. Plus, two major AI software stories: Anthropic gives Claude a built-in browser that can read, click, and type on the web, and Meta launches Muse Spark 1.1 with a new developer API, undercutting OpenAI's GPT-5.6 and Anthropic's Claude Fable 5 on price.
Subscribe for daily AI news.
AI news: 0:00 N 2:20 Boston Dynamics 2:51 Archax 01 4:05 Claude Code 5:15 Muse Spark 1.1
Japan is advancing in humanoid robotics with Highlanders' N-robot and the piloted Arcax-01, while Boston Dynamics showcases the new Atlas. Simultaneously, AI capabilities are expanding as Anthropic introduces a sandboxed browser for Claude and Meta launches MuseSpark 1.1, a multimodal reasoning model with a competitive new developer API.
Is Japan entering the humanoid robot race? Tokyo-based robotics startup Highlanders just
unveiled its Kepler V1.0 AI model that's powering its new humanoid called N. So today on AI News,
we're taking a closer look at this, plus a mecha robot that's turning science fiction
into reality. But first, Highlanders reasons that a single AI model struggles to plan and react in
Kepler accomplishes this by running on what Highlanders calls a think-slow,
react-fast architecture. A deliberative cerebrum handling vision-language reasoning and long
horizon planning at an unexpected 5Hz, paired with a reflexive cerebellum managing balance
and force control at roughly 500Hz. Plans flow down, sensor state flows back up continuously,
then at the model's core sits a 10 billion parameter world model trained on vision,
force torque, and tactile data, built specifically to predict contact dynamics
before motion happens, rather than reacting after the fact. Plus, training blends simulation,
teleoperation, and rollouts from real robots already working in the field. This is because
Highlanders robots are already collecting sensor and video data on real production floors in Japan,
feeding this loop that the company expects to generate roughly 26 million hours of targeted
real-world data in 2027 alone. And on top of this, more robots deployed means more data,
which trains a better model, which justifies deploying more robots. And running this loop
requires Highlanders Hisui infrastructure. This is their in-house cluster with an expected 640
NVIDIA B300 GPUs across 80 nodes backed up by roughly 24 petabytes of fleet sensor and video
storage. But none of this matters without somewhere for the robots to actually work,
which is why Highlanders just made it real by signing a memorandum of understanding with
Mitsubishi Motors to jointly develop and mass-produce the N-robot, targeting 1,000 units
a month by the second half of 2027 using idle space at Mitsubishi's Kyoto engine plant.
But they're facing competition from nearby Korea, as Hyundai and Boston Dynamics are also attempting
to address this human labor shortage head-on with the production of its newest Atlas robot.
In fact, Boston Dynamics just showed how far along Atlas is, as the company just released
new footage of three Atlas units performing a perfectly synchronized biking row. Mimicking the
celebration Norwegian fans made famous during the 2026 World Cup. Employees can be seen
participating in the back pulling in unison, but tell us in the comments whether you'd prefer
Atlas, or N, or else this 4.5 meter tall monstrosity called the Arcax-01, which is a
piloted mechatronic humanoid. So, a human climbs into the cockpit, built into the robot's torso,
and drives this 3.5 ton machine using joysticks, pedals, and a nine camera view feeding four
screens. And it has six 26 joints, two arms with five fingered hands, and switches between an
upright robot mode to a wheeled vehicle mode that can hit about 10 kilometers per hour,
or just over six miles per hour. In fact, the company Tsubame designed the Arcax-01 for jobs
too dangerous for people, with the CEO saying they wanted to build something unmistakably Japanese,
part mecha, part machine. But when it comes to pricing, while it started as a three million
dollar purchase, the company later added leasing to bring the cost down in 2024.
So, tell us in the comments below which humanoid form factor you'd prefer, and at what price point.
And while these robots are learning how to act in the physical world,
AI is learning how to act in the digital world too, as Anthropic just gave Claude a browser to
work with directly. Anthropic's newly added built-in sandbox browser in Claude's code desktop app
lets Claude open web pages, read them, click through, and fill out forms, including documentation,
issue trackers, design files, and any other site, all without leaving the coding environment.
On top of this, it runs on a clean profile with no saved logins, and any write action like clicking
or typing gets reviewed by safety classifiers first, even in automated mode. Claude also won't
make purchases, create accounts, or solve CAPTCHAs without explicit approval. That makes it a
different tool than the existing Claude for Chrome extension, which acts inside a user's own logged
browser sessions. This new browser is sandboxed and isolated instead, built for testing and research,
not for acting as the user. Organizations can even restrict it to an allow list of sites or
turn it off entirely. But Anthropic isn't the only company shipping new AI tools for developers,
because Meta also just changed how it sells its models entirely by introducing MuseSpark 1.1,
the second model from Meta Superintelligence Lab's division, and a major upgrade over the
original MuseSpark released in April. Alongside it, Meta opened a public preview of the Metamodel API,
the company's first real developer API for a frontier model. In fact, MuseSpark 1.1 is a
multimodal reasoning model tuned specifically for agentic work, planning, tool use, and computer use,
alongside stronger coding support, including bug fixing, large code base edits, sub-agent delegation,
and multi-step debugging loops. It also handles images, video, and documents natively in a single
call, including turning visual mock-ups directly into code. Meta.ai calls it the lab's strongest
model yet for agentic and coding tasks. And Meta is positioning their API aggressively, with reports
putting its pricing at roughly a quarter of what OpenAI and Anthropic charge for GPT 5.6 and Fable 5,
a direct play for developers currently building on either company's API. And it's a notable pivoted
strategy. Rather than compete purely on open weight availability like Llama did, Meta is now
chasing the same paid API business both rivals already run. This is different because until now,
developers wanting Meta's most capable models had two options. Either download open Llama weights
and host them yourself, or else get invited into a private preview for the original MuseSpark.
But the new API changes that by being self-serve, OpenAI SDK compatible, and including $20 in free
credits for new accounts. Furthermore, Meta says the model went through its advanced AI scaling
framework before release, with the unmitigated model reportedly crossing a high risk threshold
in chemical, biological, and cybersecurity domains. But after adding safeguards, Meta assessed residual
risk as moderate or lower, with lower jailbreak and prompt injection success rates than the original
MuseSpark. What's worth flagging though is that several independent write-ups note Meta hasn't
published a detailed model card for MuseSpark 1.1, leaving some technical specifics undocumented for
now, which is a contrast to the benchmark-heavy release post typical of OpenAI and Anthropic.
The launch also lands just two days after Meta shipped MuseImage, suggesting a broader,
faster release cadence from the Superintelligence Labs team, with MuseSpark 1.1 being expected to
eventually power Meta AI across WhatsApp, Instagram, Facebook, and the company's smartglasses.
Anyways, like and subscribe, check out this video here for more of the latest
in AI and robotics news, and thanks for watching.