Anthropic's Wildest Ten Days: IPO, Pause Call, Fable 5, Government Shutdown

Ten days. That's all it took. In the second week of June 2026, Anthropic filed to go public, co-signed a paper begging the world to slow down on AI, launched the most powerful model they'd ever built, and then watched the U.S. government yank it offline. If you work in this field, that week probably cost you some sleep. Let me walk you through it, because the sequence is almost hard to believe.

Start with June 1st. Anthropic quietly files a confidential S-1 with the SEC — that's the paperwork you file when you're heading for an IPO. The numbers are staggering. Their revenue run rate is running at forty-seven billion dollars a year, up from ten billion the year before. That's nearly five times growth in twelve months. And they're reportedly targeting a valuation north of one trillion dollars. A week later, on June 8th, OpenAI files its own S-1. So now the race to Wall Street is out in the open — two AI giants sprinting toward the public markets at the same time.

Now here's where it gets strange. Just three days after that IPO filing, on June 4th, Anthropic employees and co-founder Jack Clark publish a paper called "When AI Builds Itself." And the message is essentially: everybody, please stop.

The figures inside are what make you sit up. Claude is now writing more than eighty percent of the commits in Anthropic's own codebase. Their engineers are shipping eight times more code per quarter than they were in the years leading up to this. And there's this metric they call the "task horizon" — basically, how long a task an AI can run on its own before it needs a human. That horizon is doubling every four months. To put it in perspective: an early Opus model could handle maybe a four-minute task autonomously. A recent one can run a twelve-hour engineering task by itself. And Jack Clark puts a number on where this leads — he estimates a sixty percent chance of recursive self-improvement, AI improving itself, by 2028.

So the paper's conclusion is a call for a coordinated, conditional global pause on training frontier AI. Slow down before we lose the thread. And remember — this is the exact same week the company is deep in IPO prep.

Then June 9th arrives. Five days after the pause call, Anthropic launches Fable 5. It's their most capable public model yet — it scores 88.6 percent on SWE-bench Verified, it's tuned for enterprise and scientific work, and it's priced at fifty dollars per million output tokens. That's more than ten percent better than their previous flagship, Opus 4.8. They also quietly launch a restricted-access sibling called Mythos 5 at the same time. So the "please stop" paper is barely dry, and out comes the most powerful thing they've ever shipped.

And then the other shoe drops. June 12th, 5:21 in the evening Eastern time, a U.S. export control directive lands on Anthropic's desk. The reason: someone found a jailbreak — a known way to bypass Fable 5's safety controls — and the government is calling it a national security risk. The order is a global shutdown, effective immediately. Even foreign nationals inside the United States lose access. Anthropic offers prorated refunds. Fable 5 had been live for exactly four days.

Anthropic's response? They said the jailbreak was "relatively simple," that it worked on competitors' models like GPT-5.5 too, and that the whole shutdown was basically a misunderstanding.

So what do we actually make of all this? Let me pull out the threads that matter.

The first is credibility. When your pause paper lands on a Tuesday and your new flagship ships the following Sunday, the "safety-first" message gets muddy. And here's the honest read: Anthropic can genuinely believe both things at once — that AI is dangerous, and that they have to keep shipping to survive as a company. Both can be true. But them being true at the same time doesn't make the tension go away. It just leaves everyone watching to wonder which impulse actually steers the ship.

The second thread is the speed of regulation. Notice how the shutdown happened. It didn't come from proactive monitoring — nobody caught this ahead of time. It came four days after launch, triggered by a jailbreak somebody reported. That's reactive governance. And if task horizons really are doubling every four months, reacting after the fact is not going to keep pace. The technology moves in months; the oversight moves in headlines.

The third thread is about that pause proposal itself. It suggests trigger conditions — moments where the world should hit the brakes. But it doesn't say who decides when those triggers are hit, or how the pause gets enforced. And without a mechanism, without someone actually holding the switch, it reads less like a policy and more like a moral statement. A sincere one, maybe. But not something a government can pick up and act on.

Let me leave you with the piece I think is most worth chewing on, and it's that task horizon number. Forget the benchmark scores for a second. A model that can autonomously run a twelve-hour engineering task today — what is it doing in two years, if this trend holds? That's a fundamentally different dimension of capability. It's not about answering a question well. It's about sustained, independent action in the real world — recovering from its own errors, using tools, adapting when something surprises it. And Anthropic's own productivity data, that eight-times figure, suggests this isn't just a lab benchmark. If that number is real, then every other major lab is almost certainly seeing the same dynamics — they're just not publishing them.

So here's what I'd hold onto from this whole whirlwind. One: Anthropic is living a contradiction out loud — warning the world to slow down while racing to ship and to go public, and both sides of that are real. Two: our governance is reactive, catching problems days after launch, while the capability curve is measured in months — that gap is the real story. And three: the most important metric to watch isn't the price or the benchmark, it's how long these systems can run on their own. That's the number that changes everything else.

Fable 5 is offline as we speak. When it comes back, and under what conditions, will tell us a lot about who's really steering this — the labs, or the governments trying to keep up.