Capability raced. Trust and control set the terms.

The most powerful AI model on earth shipped on a Tuesday and was dark by Friday, ordered down by the U.S. government over the lab's objection. The hard part these two weeks was trust and control: who gets to run these systems, and who gets to pull them.

• Anthropic's Fable 5 was released, then suspended by federal export-control order three days later. The brake on a surge came from outside the labs for the first time, and governance stopped being a slide nobody acts on.
• More than $13 billion flowed into physical-world AI in two weeks. The reimagining is leaving the screen and heading for the factory floor and the lab bench.
• OpenAI declared "chat is dead" in the same stretch that Apple unveiled a conversational Siri bound for a billion iPhones this fall. The frontier and the field have never sat further apart, and the distance is a trust gap as much as a capability one.

Three things happened in the last two weeks that a strategy team has to account for. They don't reduce to a single story. What they share is a turn: the capability kept racing, and the questions that actually bit were about trust and control. Who is allowed to run these systems. Who gets to stop them. Who has to sign off before the work can run without a hand on every decision.

When the brake comes from outside the labs

Anthropic released Claude Fable 5 on June 9, the general-use version of the Mythos model I write about in the book, and the most capable system it had ever released. Its own description was vivid: work that "compressed months of engineering into days." It also arrived with governance built into it. Ask Fable 5 for something off-limits in cybersecurity, biology, or model theft, and it quietly redirects you to an older, safer model. The guardrails are part of how the work runs.

Three days later it was gone. On June 12 the Commerce Department, under Secretary Howard Lutnick, ordered Anthropic to suspend both Fable 5 and its unrestricted sibling Mythos 5 under national-security export controls, after Amazon and others reported jailbreaking Mythos in ways the government treated as a threat. The order reached so far, covering any foreign national including Anthropic's own staff, that the company pulled the models for everyone, unable to comply selectively. Anthropic disputes the order, arguing that "the finding of a narrow potential jailbreak should not be cause for recalling a commercial model deployed to hundreds of millions of people," and says it is working to restore access. As I write this, the models are still dark.

Here is why it matters past the headline. For two years governance has been the part of the agentic conversation that executives nod along to and then table. This time, it became the binding constraint on capability, enforced from two directions at once: inside the model by the lab, and outside the company by the state. The capability did not regress. Fable 5 is exactly as powerful as it was on Tuesday. What changed is who controls whether you can run it. If you had built a transformation on top of it, you would have woken up Friday with a hole where your most capable system used to be. That is the kind of slowdown the book means: the current did not weaken, an obstacle upstream stopped it for a stretch.

The reimagining leaves the screen

While one surge was being braked, another was being funded at a scale that is hard to wave off. In the same two weeks, Jeff Bezos's Prometheus raised $12 billion at a $41 billion valuation, one of the largest private financings on record, to build what it calls an artificial general engineer: software that designs and manufactures physical things like jet engines and drug compounds. Days later, Neura Robotics raised up to $1.4 billion, led by Tether with Amazon and Nvidia alongside, to push toward millions of humanoid robots by 2030.

It is tempting to read the book as a story about software agents doing knowledge work: screens, documents, inboxes. This is the stretch where that framing started to look too small. The reimagine end of the spectrum is moving into the physical world, into factories and logistics and labs, anywhere the work is not made of words. If your business makes, moves, treats, or stores anything in the physical world, which is nearly everyone outside pure software, the pressure to redesign those operations around what people and machines can do together is now landing in your building. And when the agent is a machine on a factory floor instead of a chatbot in a tab, the trust questions get heavier.

Bezos was blunt about where he thinks it goes. He said the productivity gains will mean households that need two earners today will get by on one. You can decide for yourself whether that is a promise or a warning. Either way, the capital is committed and the machines are being built.

The frontier and the field have never sat further apart

In the same two weeks that Apple finally unveiled a genuinely conversational Siri at its developer conference, rebuilt on Google's Gemini models and due this fall, OpenAI was declaring the whole category obsolete. First reported by the Financial Times, OpenAI is rebuilding ChatGPT into an agent-first "superapp" ahead of a public listing, with one insider putting it flatly: chat is dead. So the mass market is being handed a polished version of the chat era just as the frontier walks away from it.

The easy read is a capability gap. The truer one is a trust gap. Elena Verna put it well this month: most companies have "a trust problem, not a tooling problem," and what they need is "agency, not agents." Agents only pay off when the organization hands them real context and real decision room, and most organizations will not. That is the context tension from the book showing up in the field. The firms pulling ahead are quietly building the layer that makes agents useful: as David Haber observed, the default has already flipped to "assume you're being recorded," and the living record of how a company actually works is being captured whether leaders are watching or not. The uncomfortable part is the one the book names. The context you own and the influence you hold are nearly the same thing, which is what makes handing it over feel like a loss.

You can see what closing the gap looks like when it happens. a16z described the month-end close, a centuries-old accounting ritual, becoming just another day, because agents now handle almost the entire ledger continuously. That is a reimagine, a process rebuilt around what agents can do, running in production today. The firms still bolting a chatbot onto last year's close are doing a retrofit, and they are quietly ceding ground to the ones who redesigned the work.

What I'm watching next

Three things on my screen for the next issue. First, whether Fable 5 and Mythos 5 come back, and on what terms, because the precedent outlasts the model: a government that pulled one frontier system over a weekend can pull another. Second, whether the physical-AI money starts turning into deployed machines or stays a funding-round story. Third, whether the export-control net widens to other labs, which would turn a one-off into a pattern every transformation plan has to price in.