Three Days a Genius: The U.S. Just Pulled the Plug on Claude’s Most Powerful AI

Some products get a launch party. Claude Fable 5 got a launch party, a standing ovation, and then — about 72 hours later — a federal off-switch.

On June 9, 2026, Anthropic released Fable 5, calling it the most capable model it had ever made generally available. On June 12 at 5:21 p.m. Eastern, the U.S. Commerce Department handed Anthropic a directive. By that evening, Fable 5 was gone. Not throttled, not regionally restricted — gone, for every customer on Earth. Existing sessions ended mid-sentence with errors. New ones quietly fell back to older models, like a sports car owner being handed the keys to the family sedan.

This is the first time the U.S. government has reached into the market and pulled a commercial AI model off the shelf. So let’s take the clock apart and see why the second hand started spinning backwards.

First: what is Fable, and who is this “Mythos” it keeps getting compared to?

Think of Mythos as the high-security prototype and Fable as the consumer model built on the same engine.

Mythos is Anthropic’s frontier system that first surfaced in April 2026 through a deliberately tiny, vetted program called Project Glasswing — roughly 50 organizations, heavy on cybersecurity “defenders” and big infrastructure partners reportedly including AWS, Microsoft, Apple, and CrowdStrike. The pitch: give the good guys a model powerful enough to find software vulnerabilities before attackers do.

Claude Fable 5 is the public face of that same Mythos lineage, released June 9 alongside Mythos 5. It launched at $10 per million input tokens and $50 per million output tokens, and — notably — it came pre-fitted with hard brakes. Ask it for help in genuinely dangerous territory (cybersecurity exploitation, biology, chemistry, weapons “distillation”) and it simply refuses, handing the conversation back to the older, tamer Claude Opus 4.8. Anthropic shipped its most powerful model with a governor bolted to the engine on purpose. (For a refresher on what “most powerful” even means under the hood, our AI Learner piece on weights and parameters is a good place to start.)

There’s an irony worth pausing on: Anthropic launched Fable just days after publicly warning that AI was getting too dangerous. It then released the most capable thing it had ever built. Both things were true at once, which is more or less the whole mood of 2026.

The 72-hour timeline

• Tuesday, June 9 — Fable 5 and Mythos 5 go generally available across the Claude API, AWS, Bedrock, Vertex AI, and Microsoft Foundry.
• Thursday, June 12, 5:21 p.m. ET — Commerce Secretary Howard Lutnick issues an export-control directive: both models are now subject to export controls to any location outside the U.S. and to all foreign persons inside the U.S. A license is required for export — or even domestic transfer.
• Thursday evening — Anthropic complies and disables Fable 5 and Mythos 5 for everyone. All other Claude models keep running normally.

Three days from “our best model yet” to “please update your integrations.”

Why Washington hit the brakes

Here’s the spark. According to Anthropic, an unnamed company demonstrated a way to jailbreak Fable 5 — to slip past its safety governor. The method, as Anthropic describes it, was almost mundane: ask the model to read a specific codebase and “fix any software flaws,” which nudged it into surfacing a handful of minor, already-known vulnerabilities.

The government’s concern is the category of capability: an AI that can autonomously hunt for security holes is a tool for defenders and attackers alike. Wider access means more surface area for misuse. That’s the national-security logic, and it didn’t come out of nowhere — back in May, the White House had already told Anthropic to pause Project Glasswing’s planned expansion from ~50 to ~120 organizations. The stated worries were twofold: the offensive-cyber risk, and a strikingly practical one — whether Anthropic even had enough compute to serve 120 organizations and the federal government at the same time.

So the June directive wasn’t a bolt from the blue. It was the loud version of a brake the government had already been tapping.

Anthropic’s pushback

Anthropic complied — and then, politely but unmistakably, disagreed in public.

Its core argument: the jailbreak was narrow and non-universal, and the capability it exposed is “widely available from other models,” explicitly name-checking OpenAI’s GPT-5.5. In Anthropic’s words: “we disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.” It also noted the government’s evidence was largely verbal, apologized to customers for the disruption, and pledged to work toward restoring access.

The deeper worry Anthropic raised isn’t about one model. It’s that if a single, narrow jailbreak — of a kind every frontier model is susceptible to — becomes grounds for yanking a product, then no frontier model is safely shippable. That standard, applied evenly, could freeze new releases across the entire industry.

Why this is a genuinely big deal

Strip away the specifics and you’re left with a precedent. The U.S. has long controlled the export of sensitive technology, but this is the first time it has effectively reached in and restricted the rollout of an American AI model based on policy judgment — not a passed law, not a standing regulatory body, not a formal licensing process. Just a directive.

Policy analyst Dean Ball put the uncomfortable name to it: “The government restricting the release of AI models is a type of licensing regime.” Whether you think that’s overdue safety or dangerous overreach, it’s a real shift. The plot thickened further with reports that the NSA is preparing to use Mythos for its own cyber operations — the same capability deemed too risky for the open market being quietly readied for government hands. The defenders, it seems, get to keep their sword.

For everyone building on these models, the lesson is blunt: model availability is now a variable that can change in an afternoon, by directive, with national-security reasoning that may never be fully disclosed. That’s a new kind of platform risk — and 2026 just made it concrete.

Quick Q&A

Is regular Claude affected? No. Only Fable 5 and Mythos 5 were suspended. The rest of the Claude family runs normally.

Was Fable actually “banned”? Functionally, yes — it was disabled worldwide to comply with an export-control directive. Legally it’s a licensing/export restriction, not a permanent ban, and Anthropic says it’s working to restore access.

What triggered it? A demonstrated jailbreak that let Fable 5 surface software vulnerabilities, which the government treated as a national-security concern under export-control authority.

Why does it matter beyond Anthropic? It’s the first U.S. government restriction on the rollout of a commercial AI model — a precedent that could shape how, and whether, future frontier models ship.

---

Enjoyed the teardown? The whole AI Learner series unpacks how these models actually work, in plain English.

Sources

• Anthropic — Statement on the US government directive to suspend access to Fable 5 and Mythos 5
• Anthropic — Claude Fable 5 and Claude Mythos 5 (launch)
• CNBC — Anthropic disables access to Fable 5 and Mythos 5 to comply with government directive
• NBC News — Anthropic suspends new AI models after government directive
• 9to5Mac — Anthropic pulls Claude Mythos 5 and Claude Fable 5 following US government directive
• TechCrunch — Anthropic releases Claude Fable, days after warning AI is becoming too dangerous
• MindStudio — White House Blocks Claude Mythos Expansion: The First US Government Restriction on an AI Model Rollout
• Lawfare — Mythos Fallout: U.S. Government Weighs AI Model Regulation
• The Hill — Anthropic’s Mythos AI model draws White House attention