Welcome back to Level.UP, brought to you by UP.Labs.

This week, we dive into two stories. The first: Anthropic released Fable 5, and then had to disable it after a U.S. export control order. The second: NEURA Robotics raised $1.4B to build robots that learn from every facility they work in, which could now make your factory floor part of someone else's business model.

Plus: Bezos is funding an "Artificial General Engineer,” AI powers the World Cup, and more from the world of physical AI.

Think someone else needs this? Forward it to a friend or colleague navigating the same terrain.

Three days after Anthropic released Claude Fable 5 — the most capable AI model ever made available to the public — the US government pulled it.

On the evening of June 12, an export-control directive citing national-security authorities ordered Anthropic to suspend access to both Fable 5 and its restricted sibling, Mythos 5, for any foreign national — inside or outside the country, including Anthropic's own foreign-national employees. Because the company can't filter out American users in real time, it disabled both models for everyone. Anthropic's other models, including Claude Opus 4.8, stayed online.

The stated trigger was a verbal warning about a "narrow, non-universal jailbreak," essentially prompting the model to read a codebase and flag security flaws — a capability Anthropic notes is common to competing models and used daily by cybersecurity professionals.

Anthropic disagrees that a narrow jailbreak warrants recalling a model deployed to hundreds of millions, and flew staff to Washington to resolve it. As of mid-June, the dispute is unresolved — making it the first government-forced takedown of a publicly deployed frontier model.

While this is a developing situation, the resolution matters less than what it has already revealed.

For any operator, the lesson is that frontier AI is no longer just a procurement decision. It's now subject to a switch that neither you nor your vendor may fully control. If your operation builds a critical workflow on a single frontier model, you've inadvertently taken on a dependency that could be suspended by a directive you may never see coming. The defensive move is unglamorous and familiar: a fallback model, a portable prompt layer, and the ability to fail over without halting the line.

For non-US operators, the signal is sharper. A model your business relied on was switched off, in part, because of where some of your people hold passports. That is a reason to question how much of your operation should sit on American AI infrastructure at all — and many will rationally start hedging toward models they can count on staying on.

Step back, and the move looks less like safety enforcement and more like statecraft. When an executive branch can decide who gets access to frontier capability, that capability stops being a product and becomes an instrument, potentially impacting global confidence in American AI institutions. This issue isn't uniquely American, though — Beijing pulls strategically important companies back into its orbit too, like the recent Meta-Manus merger block.

The world's two AI powers are both treating models as national assets to be granted and withdrawn. We are watching frontier AI become an arm of foreign policy in real time.

What that means in practice is simple: the operators who plan for a world where the best model can vanish overnight, rather than one where it's simply available, will be the ones still running when the next switch flips.

NEURA Robotics, a German cognitive-robotics company, announced a Series C of up to $1.4B. The investor list includes supply chain giants like Qualcomm, Amazon, and NVIDIA in compute and cloud; Bosch and Schaeffler from Europe's industrial core; the European Investment Bank; and Tether backing a machine payments layer.

NEURA reports an order book above $1B and plans to scale production toward "multi-million robots by 2030.” The capital targets the Neuraverse — an open ecosystem where deployed robots share skills and learning across sites — and “NEURA Gyms,” large-scale physical training environments built to generate real-world robotics data at volume.

"Many believed globally relevant AI infrastructure companies could only emerge from Silicon Valley," said founder David Reger. "The next generation of AI leaders can emerge anywhere in the world where there is enough vision, engineering talent, and execution speed.”

Last week, we argued that physical-industry operators are crucial inputs to the AI economy. The labs can build intelligence, but they don’t own the factories, supply chains, or decades of operational data that intelligence needs to be useful in the real world.

This funding round supports that thesis. NEURA's robots are designed to learn from every site they work in and share that learning across the whole network. That's the business model: each deployment makes every robot smarter, which sells more deployments.

But notice what feeds it. When one of these robots works in your facility, it learns from your hardest problems, your know-how, your mistakes — knowledge your company spent years and real money acquiring. 

That knowledge is precisely the asset these platforms are raising billions to collect. NEURA is building entire training facilities to generate what your operation produces as a byproduct. So before you sign a pilot, be sure to consider the value of what you're contributing. What does the robot's maker keep? Can skills learned on your floor end up helping your competitors who use the same robots? And what do you get in return for teaching the system?

None of this means avoiding these platforms. The skill is knowing which is which: routine tasks where you benefit from everyone's learning, and proprietary processes that shouldn't go anywhere near a shared system.

Companies learned this lesson too late with cloud software, when they discovered their data had become someone else's product. This time, the lesson is available in advance. The robot is the part of the deal you can see. The learning loop is the value. Don’t forget that you're on both sides of the transaction.

Ready to work smarter? Here are the tools we're tracking this week:

Bezos Is Funding An "Artificial General Engineer." Prometheus, co-founded by Jeff Bezos and Verily's Vik Bajaj, raised a $12B Series B at a $41B valuation to automate the design of complex physical systems, from jet engines to drug compounds. That's a 150-person company valued at roughly $270M per employee, with most of the capital going to compute. The signal: frontier money is now aimed squarely at engineering work itself — the exact talent physical operators already can't hire enough of. → Read more

In China, Car Brains Are Becoming Robot Brains. Autonomous-driving firm MINIEYE and robotics maker LDROBOT are merging their technology and data infrastructure into one shared Physical AI foundation — feeding delivery vehicles, service robots, and eldercare machines from the same learning loop, with pilots in Shenzhen and Changsha. The thread: robots are interchangeable; the pooled data is the product. → Read more

The World Cup Is A Physical AI Showcase. This World Cup, every player gets a 30-second 3D body scan, and the dimensionally accurate avatar it produces feeds the semi-automated offside replays. Meanwhile, referees across all 104 matches wear head-mounted RefCams — Lenovo's AI stabilization cuts motion blur by up to 50%, streamed over private 5G — turning the official's own point of view into a new data feed. The takeaway: the hard, defensible layer in physical AI is the sensing and capture infrastructure that grounds it in real bodies, real sightlines, real latency. Whoever owns that layer owns the inputs that everyone else's intelligence runs on. → Read more