The frontier model that vanished on a Friday night.

The letter arrived at 5:21pm on a Friday, the hour when most of corporate America has already started its weekend. According to Anthropic, that was the moment a government order landed in its inbox, and the timestamp has become part of the story. Al Jazeera reports the company received the directive at 5:21pm local time, and that the letter did not spell out the government's specific security concern in any detail.

What it did do was unambiguous. As Anthropic described it, the U.S. government, citing national security authorities, issued an "export control directive" ordering the company to suspend all access to its two newest and most powerful models, Claude Fable 5 and Claude Mythos 5, for any foreign national, whether inside or outside the United States. The order reached all the way into Anthropic's own staff: foreign-national employees were to be cut off too.^[1][2]

Here is the wrinkle that turned a targeted order into a global blackout. Anthropic says it had no clean way to fence off only foreign nationals, so to comply, it pulled the plug for everyone. A paying customer in Ohio lost the model at the same instant as anyone abroad. A frontier system that had been public for only a handful of days simply vanished from the menu.^[1]

01 · What was switched off

Not a toy. The most capable model the public had ever touched.

To understand the stakes, you have to understand the tier that was yanked. Reporting from XDA Developers lays out Anthropic's ladder: Haiku is the fast worker, Sonnet the everyday balance, Opus the heavy lifter. Above all of them sits something newer and stranger: the Mythos class.

Until this month, Mythos models were not for the public at all. They were restricted to U.S. government cybersecurity partners through a program reportedly called Project Glasswing, because Anthropic considered them too dangerous to release broadly. Claude Fable 5 was the first Mythos-class model offered to anyone with a subscription, and per XDA, it was the same underlying model as Mythos 5, the only difference being a layer of safety classifiers screening cybersecurity, biology, chemistry and model-distillation queries.^[3]

On the benchmarks it was a monster. Per XDA, Fable 5 scored 80.3% on SWE-Bench Pro, eleven points clear of GPT-5.5, and 13.4% on Cognition's FrontierCode Diamond, more than double GPT-5.5's 6.3%.^[3] The model that got pulled was not a curiosity. It was, by these numbers, the most capable coding model the public had ever briefly been allowed to touch. And it disappeared in an evening.

02 · The reason no one will explain

A threat the public has never been shown.

So why did the hammer come down? On the official record, almost nothing. Al Jazeera notes the letter itself did not explain the government's specific concern; The Conversation reports the government has not publicly stated a reason at all.^[2][4]

What we have instead is Anthropic's account, and the attribution matters. According to Anthropic, the government had become aware of a jailbreak: a way to slip past the safeguards that are supposed to reroute dangerous requests to a weaker model, potentially to extract information useful for cyberattacks. But the company is plainly skeptical of how serious it is. Decrypt reports Anthropic says the government has so far offered only verbal evidence of a "narrow, non-universal" jailbreak (essentially, asking the model to read a codebase and fix its flaws) and that this capability is already widely available from competing models, GPT-5.5 included.^[4][5]

03 · The fingers, and where they point

Three pressures, carefully attributed.

This is where it stops being about a model and starts being about institutions leaning on one another. Several of these are claims made by interested parties, so they are worth naming precisely.

• The Pentagon. The Conversation reports the relationship had been souring: the dispute sharpened when Anthropic declined to let the Pentagon use its models for domestic surveillance and fully autonomous weapons, after which the Department of Defense threatened to designate Anthropic a "supply chain risk," a label that would have forced military contractors to cut ties.^[4]
• Amazon. The sharpest finger is Anthropic's own. The Conversation reports Anthropic says the research behind the directive appears to have been produced by engineers at Amazon, at once a rival in the model race and one of its largest investors. Amazon's side of that is not established in the reporting.^[4]
• The jailbreak underground. The guardrails were also hammered from outside: within 48 hours of release, a researcher using the handle "Pliny the Liberator" published what they identified as Fable 5's full system prompt to X and GitHub. Anthropic concedes the broader point: "perfect jailbreak resistance is not achievable for any current model provider."^[4]

And it would be too tidy to cast Anthropic purely as a victim. It chose to release a model it had previously deemed too dangerous for the public, wrapped only in classifiers, and it chose a total blackout for paying users as the way to comply with a narrow order. Those were decisions, not weather.

04 · Why aërgap cares

You can't get switched off from a system you own.

Strip away the personalities and one fact remains, and it is the one that should travel furthest. A government can now, on a Friday evening, with a letter that explains nothing in public, reach into a U.S. company and switch off its best product, not just for adversaries abroad, but for its own engineers and for ordinary domestic customers who did nothing wrong. If you had built your product on Fable 5 that week, your roadmap went dark at 5:21pm on someone else's say-so.

Everything we build at aërgap comes from one stubborn position: you should be able to own the tools that run your company, instead of renting them by the seat from a vendor who holds the off switch. When the model runs on hardware you control, as we wrote about open-weights models like Gemma 4, the failure modes in this story simply don't exist:

• No remote off switch. A directive to a vendor can't reach a box you own
• No model rug-pull. Open weights can't be deprecated out from under you
• Works air-gapped. Run it on a network with no internet at all
• Nothing leaves your tenant. Prompts never touch a third-party API
• Your roadmap is yours. No one else schedules your outages
• Auditable end to end. You hold the weights, the logs, the prompts
• No per-token meter. Once deployed, inference is just electricity
• Compliance by construction. The data stays where your server is

The honest caveat is the same one we always give: a model you run yourself may not be the very top of the frontier on the hardest tasks, and someone has to provision and maintain it. That is the trade every self-hosted system asks for: rent a capability, or own a system. For most of the last three years the cloud API was the only answer that worked for AI. This Friday was a reminder of what the easy answer actually costs: a dependency that someone else can revoke before the weekend.

The companies that move first will be the ones who already did this math for the rest of their stack. If your helpdesk, your ERP and your data already live on infrastructure you own, the model is just the last piece to bring home, and the one whose absence, it turns out, can arrive by letter.