Before I left the office Friday night, I spent an hour and a half writing up the design for a tool I’d wanted: a way for any of my coding sessions — Claude Code or Codex — to send or check e-mail when I asked. Not the code — just the plan, in plain English. When it was all on the page, I turned on Remote Control in my Claude Fable 5 session and told it to build the whole thing end to end while I drove home.
By the time I’d covered the ten miles from the office to a diner near home and looked at my phone, Fable 5 was done — the entire implementation, carried out on its own over the internet while I was on the road. A moment later the screen returned an error: the model claude-fable-5 was unavailable.
It wasn’t a glitch. It was the federal government.
That last job wasn’t a one-off, either. From Tuesday’s release to Friday night I used Fable 5 as much as time allowed — about 1.6 million tokens of output, drawn from roughly 225 million tokens of code and context, adding up to just under $440 of work at Anthropic’s published API rates. Free on my Max 20x subscription, all of it work I’d have paid for. Four days, and it was over.
What the government did
The directive reached Anthropic at 5:21 p.m. Eastern on Friday. By the company’s account it was an export-control order citing national security authorities, suspending all access to Fable 5 and its unrestricted sibling, Mythos 5, for any foreign national anywhere — including Anthropic’s own non-citizen employees. The New York Times reported it came from the Commerce Department, with no word on how long it would last.
There’s no clean way to block only foreigners from a model that serves everyone, so the effect was a full shutoff: both models went dark for all customers. Opus 4.8 and the older models were untouched.
The basis is the strange part. Anthropic says the letter gave no specifics, but its understanding is that the government believes someone found a way to “jailbreak” Fable — a technique that, as the company describes it, essentially amounts to asking the model to read a codebase and fix its flaws. Anthropic reviewed the demonstration and said it surfaced a few minor, already-known vulnerabilities that other public models, OpenAI’s GPT-5.5 among them, turn up with no trick at all.
The company complied, apologized to its customers, and then said plainly that it disagrees, calling the episode a misunderstanding. Recalling a commercial model used by hundreds of millions over a narrow jailbreak, it warned, would halt every new model release if the standard were applied industry-wide.
Not the first time
This is the second time in three months the administration has gone after Anthropic. In March, after talks over military and intelligence use broke down, the Pentagon labeled the company an “unacceptable supply chain risk.” Friday’s order went further — written broadly enough that, as the Times noted, it could in theory bar Anthropic’s own Canadian or British engineers from working on the models they build.
What stood out was where the criticism came from. Dean Ball, who advised this same White House on AI policy, wrote “I have no words” and called the decision “baffling,” and former cybersecurity and technology-policy officials told the Times they were surprised, calling it highly unusual. It’s hard to square with the same administration’s recent move to let advanced AI chips be sold to China. And when OpenAI restricted its own cyber model this spring, it did so on its own terms — no government order required.
Why this one is personal
I should be honest about where I stand. Back in February and March, when the Pentagon first moved against Anthropic, I wrote that the administration had lost my support over its treatment of the company. I haven’t changed my mind, and Friday’s notice brought the anger back.
Anthropic is far more than a vendor to me, and a threat to it feels personal in a way that wouldn’t be true for any other company on earth. Over the past year Claude has gone from a curiosity to what I reach for first on all the problems — the kind that can eat a team’s day but trace to a root cause in fifteen minutes. My access to Fable was temporary anyway — on my plan only through June 22, after which keeping it meant paying by the token. Losing it nine days early should be a shrug. It isn’t. If a cloud provider or a phone-system vendor hit a regulatory wall, I’d note it and move on. This one I feel.
The honest caveat: None of us have seen the classified case, and maybe the government is looking at something we can’t. But I’m past giving Washington the benefit of the doubt, and on the facts Anthropic and the Times have put on the record — a “jailbreak” that comes down to fix the bugs in this code, turning up flaws competing models find on their own — switching off the most capable model the public has ever been handed, for everyone, doesn’t add up.
It will probably come back; Anthropic says it’s working to restore access. But the model was never the point. The point is that the government can reach into a running American product, switch it off overnight on a security theory it won’t fully explain, and leave the company no move but to comply. That should hold the attention of people who will never type a word into Fable 5. It was nice while it lasted; the precedent is the part I’d watch.