News

The US Government Just Blocked Anthropic's Fable 5: Here's What Happened

The US ordered Anthropic to cut off Fable 5 for foreign nationals, so it pulled its most powerful model offline worldwide. What happened, and why it matters.

MesmerTools•June 13, 2026•14 min

Anthropic shipped the most capable model it has ever built on a Tuesday. By Friday night the US Commerce Department had ordered it switched off for anyone who isn’t a US citizen. There’s no dial that sorts API traffic by passport, so Anthropic did the only thing it could and pulled Claude Fable 5 and Claude Mythos 5 offline for the entire planet. Here is what happened, why it happened, and the lesson every builder should take from it.

Quick Take

The short version: On Friday, June 12, 2026, Commerce told Anthropic that foreign nationals could no longer touch Fable 5 or Mythos 5. Anthropic can’t filter users by citizenship in real time, so it killed both models globally. Paying enterprises, its own staff, everyone. Every other Claude model still works.

People are calling this the first time US export controls have been pointed at a deployed American AI model instead of chips. Anthropic is complying under protest and says it wants the models back online.

What actually happened

At 5:21 PM Eastern on Friday, a letter from Commerce Secretary Howard Lutnick landed in Dario Amodei’s inbox. Citing export-control authority, it told Anthropic to cut off Fable 5 and Mythos 5 for every foreign national, including the non-citizen engineers who work at Anthropic inside the United States.

Here’s the problem with that order: nothing in a normal API stack checks your citizenship before answering a prompt. Anthropic had two choices. Risk being out of compliance, or turn the models off for everyone. It chose the kill switch. Inside a few hours the best model on the market went dark worldwide. Opus 4.8, Sonnet 4.6, and Haiku 4.5 kept running.

Per Axios, the trigger was an unnamed company claiming it had jailbroken Mythos 5 to reach its cyber-exploitation skills, which spooked the administration. Anthropic doesn’t think the technique is anything special, and we’ll get to why.

The timeline

June 9, 2026

Fable 5 and Mythos 5 launch

Anthropic releases Fable 5 as the public, safeguard-wrapped sibling of the restricted Mythos 5. It costs $10 per million input tokens and $50 per million output tokens, and it’s pitched as the most capable model anyone can buy.

June 10, 2026

Amodei publishes an AI-policy essay

Dario Amodei posts “Policy on the AI Exponential,” arguing for mandatory pre-deployment testing of frontier models and government authority to block releases that cross a risk line. Forty-eight hours of foreshadowing, as it turns out.

June 12, 2026, 5:21 PM ET

The Commerce letter arrives

Secretary Lutnick orders foreign-national access to both models suspended on national-security grounds.

June 12 to 13, 2026

Both models go dark, worldwide

Unable to gate by nationality, Anthropic disables Fable 5 and Mythos 5 for every user on Earth, then publishes a statement that disputes the rationale while pledging to comply and restore access.

What Fable 5 and Mythos 5 actually are

Mythos 5 is the frontier research model, locked to pre-approved partners. Fable 5 is the one the rest of us can rent: the same underlying architecture, wrapped in classifier-based safeguards that block responses in sensitive areas (cybersecurity, bio and chem, model distillation) and fall back to an older model when a request crosses the line.

The capability that got the government’s attention is the family’s knack for finding and exploiting software vulnerabilities. That’s the textbook dual-use problem. The exact skill that helps a defender patch a zero-day helps an attacker find one. Once a model is genuinely good at reading code for flaws, “who gets to use it” stops being a purely commercial question.

The jailbreak at the center of it

Anthropic calls the exploit a “narrow, non-universal jailbreak.” In plain terms: ask the model to read a specific codebase and fix its security flaws. The company’s argument is that this is table stakes for any frontier model (it pointed straight at a competitor that does the same thing) and that defenders use it every day to keep systems safe. Enforce this standard across the industry, Anthropic warned, and you “essentially halt all new model deployments for all frontier model providers.”

Anthropic’s response: comply, but contest

Anthropic complied on the spot and argued in public. Its statement said the letter “did not provide specific details of its national security concern” and slammed the process for skipping the basics: no documented evidence, no appeal window, no risk tiering. The company says it’s working to bring access back “as soon as possible.”

The people who got hurt fastest were builders. Fable 5 had been generally available for exactly three days. Teams that wired it into production over the weekend, and started paying for the privilege, watched the dependency evaporate on a Friday afternoon with zero notice and no migration path.

The irony nobody let slide

Two days before the order showed up, Amodei had publicly argued that the government should be able to block dangerous frontier models. Then the government blocked his. TechCrunch ran with the line that Anthropic’s safety warnings “may have just backfired.” To be fair to him, Amodei asked for capability-based blocking with guardrails against political misuse, and what landed looks to a lot of people like a political move wearing a national-security costume. That gap is the entire argument.

This isn’t the first US move against Anthropic

Earlier in 2026 the fight was about defense, not exports. After a dispute over the Pentagon’s demand to use Claude for “all lawful military purposes” without Anthropic’s guardrails, the administration ordered federal agencies to stop using Anthropic’s tech in late February, and the Defense Secretary tagged the company a “supply chain risk.” That label normally gets reserved for foreign-adversary hardware.

A federal judge later froze that designation, calling it likely “contrary to law and arbitrary and capricious,” though the Pentagon held that the restriction stood for its own purposes pending appeal. The takeaway for context: a US government that’s willing to aim national-security tools at a domestic AI lab is now a pattern, not a fluke.

The precedents for blocking an AI tool

Governments have throttled AI tools before. The catch is that they almost always went after foreign ones, and almost always with device bans rather than a global kill switch. Fable 5 is the odd one out because the target is a US company’s own model. Here’s how it stacks up against the recent playbook.

Action	Target	Mechanism	Scope
Fable 5 / Mythos 5 (Jun 2026)	US frontier model	Export-control directive	Foreign nationals, then global shutdown
DeepSeek bans (2025)	Chinese model	Government-device bans (14+ states, federal agencies)	Gov devices only
TikTok (2024 to 26)	Chinese-owned app	Divest-or-ban law upheld by SCOTUS	Resolved via ownership restructure
ChatGPT in Italy (2023)	US model	GDPR emergency block	One country, about a month
AI chip export controls	Hardware	BIS Entity List plus compute thresholds	Exports to China / Tier-3

One more thing about the backdrop, because it matters. The Biden-era “AI Diffusion” rule that would have put export controls on model weights above a compute threshold was rescinded in May 2025 and never replaced. The Senate killed a proposed 10-year moratorium on state AI laws in mid-2025, so California (SB 53) and Texas (TRAIGA) went ahead with their own frontier rules. Translation: the federal framework here is improvised, which is a big part of why this order is so contested.

How the internet reacted

X, Hacker News, Slashdot, and every builder channel lit up, and the room was split. Developers were furious about losing a paid model overnight. Policy people, including some who back export controls, picked apart the process. European officials treated it as a wake-up call for AI sovereignty. The quotes below are as reported by the outlets we cite, and we stuck to the ones that got picked up widely.

“This means you should expect to have to prove your citizenship to use Anthropic models.”

Dean Ball · AI policy, ex-White House · X

“This week the most advanced AI model on the planet got switched off by a foreign government. This isn't an AI story. It's the story of every industry we used to lead.”

Al Carns MP · UK, former armed forces minister

“Washington's decision to cut access to Anthropic's most powerful models should serve as a wake-up call.”

Bruno Retailleau · French interior minister

“Just tried Fable. It burned 1.3M tokens in 7 minutes. That's $160 per hour.”

Per Borgen · Scrimba CEO · X

The themes repeated themselves. Developers warning each other to never put a brand-new model on the critical path. Open-weight and local-model people treating the shutdown as their closing argument. And a steady drumbeat of “why only Anthropic,” given that comparable capabilities exist at other labs. The strongest rebuttal is blunt: if a model is genuinely good at weaponizable cyber-exploitation, a government erring on the side of caution with incomplete evidence is doing its job.

The fairest read

Both things are true at once. The cyber capabilities of a frontier model are a real national-security interest, and Anthropic spent months telling everyone that Mythos-class models are dangerous enough to need unusual controls. And a directive with no public evidence, no appeal, and no clear legal footing, applied to one company while equivalents stay online, is exactly the kind of move that burns trust in the rules. You can hold both without flinching.

What this means for builders

The lesson is not “avoid Anthropic.” It’s that any single proprietary frontier model is a single point of failure, and the failure mode just got worse. A model used to disappear because of an outage or a price hike. Now it can vanish by government order, overnight, with no notice and no off-ramp. If your product breaks when one model does, that’s a decision you made, and it’s worth revisiting this week.

Boring hygiene that aged beautifully over the weekend: keep fallback logic in the stack, log the exact model ID behind every output, hide your provider behind a thin interface, and never make a two-day-old model the load-bearing wall of anything in production. Resilience is cheap. Finding out you didn’t have it is not.

Build on tools that don’t bet everything on one model

The tools in the MesmerTools directory are picked for people who want to ship, not babysit one vendor’s mood. Our own products are deliberately model-flexible. bestphotoAI handles image and video generation, and Framecall handles AI motion videos, so one model going dark doesn’t take your workflow down with it.

Browse the directory

Frequently Asked Questions

Is Fable 5 banned everywhere now?+

Functionally, yes, for now. The US directive only restricted foreign-national access, but because Anthropic can't filter users by citizenship in real time, it disabled Fable 5 and Mythos 5 globally for all users. The company says it's working to restore access.

Can US citizens still use Fable 5?+

Not as of this writing. The order targeted foreign nationals, but the only way Anthropic could comply was a global shutdown, so US users lost access too. A real fix would need some kind of access gating that doesn't exist yet.

Are other Claude models affected?+

No. The directive named only Fable 5 and Mythos 5. Claude Opus 4.8, Sonnet 4.6, and Haiku 4.5 stayed up the whole time.

Why did the government step in?+

Reporting points to a claim that someone jailbroke Mythos 5 to reach its cyber-exploitation capabilities. Anthropic calls the technique a narrow, non-universal jailbreak that other models can do and that security defenders use routinely.

Is this really the first US block of an American AI model?+

It's widely described as the first use of export-control authority against a deployed US commercial AI model, as opposed to chips or hardware. The earlier 2026 Pentagon dispute kept Anthropic out of federal use, but that was a procurement and supply-chain action, not an export control.

What should I do if my product depended on Fable 5?+

Fall back to Opus 4.8 or another comparable model, log which model produced what, and put a provider abstraction in place so the next surprise is a config change instead of a rewrite. Treat brand-new models as experimental until they've proven stable.

Sources & further reading

This is a developing story reported on June 13, 2026. Details, including whether and when access comes back, may change. Social-media quotes are reproduced as cited by the outlets above and have not been independently verified against the original posts.

Back to all articles