Intelligence Has an Off Switch. It's Not in Your House.
Anthropic launched Fable 5, the most capable model they'd ever built. Three days later, the US ordered it shut down for the entire planet. What the blackout reveals about the new castes of intelligence and why you need a model-agnostic architecture.
If you've seen Elysium, you remember the world it sets up. An orbital station where the elite live, with clean air and machines that cure any disease in seconds. And below, on a ruined Earth, the rest of humanity looks up dreaming of that life. The technology that changes everything exists. It's just not for everyone. It's one orbit away, behind a system that decides who gets to go up and who doesn't.
The same old class-struggle story, but taken to a fairly particular dystopian extreme. And something similar happened this past weekend. For a few days we had the feeling of living in the orbital station, where AI was giving ordinary humans superpowers. But that feeling didn't last long, because one phone call sent us back to Earth.
On June 9th, Anthropic launched Fable 5, the most capable model they had ever built. Records across nearly every industry benchmark. On June 12th, at 5:21 p.m. Eastern time, they received a letter from the US Department of Commerce. By 10 p.m., the model was shut down. Not for some. For everyone, worldwide, including customers who were paying to use it and Anthropic's own employees who aren't US citizens.
Those who tried Fable got a glimpse of where we're headed. And it's hard not to think of Elysium and take it as a given that superintelligence will only be for some.
On cost: Fable 5 would have cost twice as much as its predecessor, Opus 4.8, and was only going to be available via API, which already makes it less accessible to the average user. Putting it as an option in the chat was simply a "tasting" of what was coming, meant to build future demand.
And not just that. It became clear that, beyond the economic factor, your nationality or any other factor could in the future determine whether that "superior" entity grants you access to the holy grail of "intelligence."
What happened, without the soap opera
A whole circus of theories sprang up around this, so it's worth standing firm and separating the fact from the mythology and the story's own fable-making.
The fact, as reported by the Wall Street Journal: the move was triggered by Andy Jassy, Amazon's CEO, in conversations with government officials, including the Treasury Secretary. Amazon researchers had used a series of prompts to jailbreak Fable 5 into revealing information that was supposed to be blocked, about software vulnerabilities. Jassy took it to Washington. The government demanded the flaw be fixed or the model taken down, and chose the more drastic path: banning access for any foreign user, which led Anthropic to shut it down for the entire planet.
Worth clarifying: according to security experts cited by the Journal itself, the information extracted was basic, and other public models already provided it. It wasn't a weapon. That makes the reaction even more revealing. This wasn't really about technical severity, it was about the power to decide who gets to pull the plug on what.
And here's a second detail your judgment needs to register. Amazon is an investor in Anthropic and supplies the hardware in its datacenters that runs its models. It's partner and competitor at once. The one who pushed the button wasn't a distant enemy. It was someone sitting at the same table, straight out of Game of Thrones.
There's a lot of speculation about the ultimate why. Maybe it was to stall Anthropic's IPO. Maybe the government wanted control over the most valuable company of the moment. Maybe it was payback for a prior clash with the Pentagon. All plausible, none confirmed.
The only confirmed thing is the only thing we need to take from all this: a frontier model can be shut down globally, overnight, by a political decision you don't control and don't see coming.
The castes of intelligence
A few weeks ago I wrote about the two ERP castes: those who marry the vendor and migrate to the cloud tied to its roadmap, and those who stay on stable systems and solve AI on their own terms, without handing over the keys. The Fable 5 blackout is that same idea, raised a level, and far more serious.
It's one thing to depend on a vendor, where at least there are contracts and support timelines. It's another to build your business on a model a government can declare a controlled export weapon and shut down in ninety minutes. Without warning. If your product was a wrapper around a frontier model, like so many SaaS companies, it can stop existing with one email. Not because of any mistake you made, but because two or more powers were testing each other's resolve and you happened to be standing in the middle.
And here the Elysium metaphor stops being a metaphor. A new stratification is taking shape, and it isn't between those who use AI and those who don't. That's a subplot nobody cares about anymore, the naive version. The real divide is structural access.
At the top are those who build the frontier models: the United States and China. They set the rules, the prices, and, as we saw, who gets let in. One rung below, allies with some negotiating power. Europe lives here, and it's uncomfortable: it has the UK involved in testing these models, but not a single European company among the frontier builders. Mistral is the closest the continent has, and it plays several rungs below. Europe drafts the world's best regulations for a technology it doesn't produce. It legislates the menu of a restaurant where it doesn't cook. And further below still, everyone else. Latin America, where dependence is nearly total and hope is in short supply.
When this hits your ERP
A few weeks ago I wrote on LinkedIn about something that now fits perfectly into this story. At Sapphire 2026, SAP announced Claude, Anthropic's model, as the main reasoning layer for Joule and its entire Business AI platform. More than 200 agents and fifty-odd assistants running on a single provider's engine, embedded in finance, supply chain, HR, and procurement across hundreds of thousands of companies.
Now reread the Fable story with that information layered on top.
Anthropic's frontier model got shut down by a government order. What happens to your company's critical processes the day the engine reasoning over them doubles in price overnight, as was about to happen with Fable? Or the day someone, for whatever reason, decides to turn it off? I'm not saying Joule is going to fall over tomorrow. I'm saying tying your business processes to the roadmap, pricing, and availability of a single frontier provider is an architectural decision with consequences your procurement team should be evaluating now, not after rollout.
And there's a practical aggravating factor anyone who's ever budgeted for AI knows well. Cost forecasting with these models is already hard today, with prices that move and consumption that's half per-seat and half actual usage. If on top of that the price can double or access can be cut off for geopolitical reasons, budgeting stops being a spreadsheet and becomes a bet.
What we have on this side of the third world
In February, Chile launched Latam-GPT, the first open language model built from and for Latin America. It's led by CENIA, Chile's AI center, with support from several countries, Uruguay among them.
It's a bit unfair to compare this roughly half-a-million-dollar model to the frontier ones, where billions are flowing into training and infrastructure. You don't need to be an expert to know it has zero chance of competing with the giants. And to add irony to insult, its first version was trained on Amazon's cloud. The same one that pushed the button to put Fable 5 to sleep.
But that's exactly the point, and I want you to separate the emotion from the analysis. Latam-GPT doesn't compete with Fable 5, nor does it try to. It's a sovereignty insurance policy still in development. It's the difference between having a vegetable garden in your backyard and trying to compete with the supermarket chain. The garden won't make you rich. But the day they shut the supermarket down over an order you don't control, it'll at least feed you for a while.
A lot of people will say "but you've got the Chinese models, they're basically free." It's naive to think the Chinese are doing this out of charity. And there's nothing better than having something of your own, even if it isn't frontier-grade. All things considered, I think it's a good sign that something like this is being built around these parts.
The Operator's response
The lesson from Fable 5 isn't "don't use frontier models" because they can be taken away from you. That would be stupid, they're the best there is. The lesson is don't build on a single point of failure you don't control. And from that come two principles:
The first is to become model-agnostic. Sensible architecture treats models as interchangeable commodities. Your system shouldn't be asking "how do I talk to Fable 5?" but rather "what's the best model for this task, and how do I swap it out without rewriting everything if it disappears tomorrow?" The value isn't in the model, it's in the layer that orchestrates it, the one that lets you change engines without changing the car.
The second is sovereignty where it actually matters. Not everything you do needs the most powerful model on the planet. That meme about centering a div with Opus 4.8 points exactly at this: wasting resources and money on tasks smaller models can solve. A good chunk of daily work, especially anything touching sensitive data, should run on local models, on your own hardware, with runtimes that today are accessible to anyone with a decent machine. It's not about saving money, although you do. It's because if they run on your on-premise infrastructure, no letter from any country's Department of anything can shut them down. For healthcare, legal, or finance, which by law often can't send data to US servers, this isn't a preference anymore, it's the only viable door.
A formula that works is:
Frontier model for raw power, local model for whatever requires guaranteed privacy and continuity. That combination isn't paranoia, it's architecture.
The underlying question
Frontier intelligence is going to stay concentrated in two countries and a handful of companies, treated as a strategic asset on par with advanced chips or nuclear material. What happened with Fable is simply the start of a controlled cold war, where Elysium could become more of a reality than just another Hollywood dystopia. And that possible reality isn't something you're going to change from your desk in Montevideo, Madrid, or Bogotá.
What you do control is your exposure to it and how you manage it, assuming you even have access. You can build your work like a tenant who can be evicted by a letter that arrives on a Friday at 5:21 p.m. Or you can build it knowing from day one that the engine is borrowed, and design so you can swap it out.
Intelligence has become the new electricity for many areas of life, and that's despite being far from mature or fully autonomous. This past weekend we learned it has an off switch, that it's not in your house, and that whoever controls it isn't going to ask your permission in the slightest.
Whoever understands this designs for autonomy. Whoever doesn't, prays they don't get cut off. Let's hope the next frontier model launches aren't a fable, like Anthropic's latest, which was even named that. Every fable is brief by nature, and like every fable we should find its moral. Deep down, I hope the little animals in this story aren't an allusion to ourselves.
Pablo / En Producción
References:
- Ramkumar, A. & McMillan, R. "Amazon CEO's Talks With U.S. Officials Triggered Crackdown on Anthropic Models." The Wall Street Journal, June 13, 2026.
- Anthropic. "Statement on the US government directive to suspend access to Fable 5 and Mythos 5." June 12, 2026.
- SAP News Center. "SAP and Anthropic to bring Claude to SAP Business AI Platform." Sapphire 2026, May 12, 2026.
- CENIA. "Latam-GPT: la primera IA regional abierta creada con datos latinoamericanos." February 2026. Statements from Gabriel Boric and Álvaro Soto.
- Euronews. "What is Latam-GPT." February 2026. Includes warning from Alejandro Barros (Universidad de Chile).
Newsletter
In Production
Every Tuesday: what is actually happening with AI inside large enterprises, written by someone implementing it there.