The Fable Affair:
When a government
killed the world's best AI.

On the morning of June 9, 2026, Anthropic released what it called the most powerful AI model ever made available to the general public. By the evening of June 12, it was gone. Not deprecated, not replaced — switched off globally by a letter from a US government official, taking effect in 90 minutes. Hundreds of millions of users, developers mid-task, enterprise pipelines in flight — all dark, simultaneously, because of a single export control directive.

It stayed down for 19 days. It came back on July 1. And nothing about how the world thinks about building on AI infrastructure will be the same again.

What Fable 5 actually was

To understand why the shutdown mattered, you need to understand what Fable 5 was. Not in marketing terms — in real terms.

Stripe used it to refactor a 50 million line Ruby codebase in a single day. A project their engineering team estimated at over two months of work. That is not a benchmark number. That is a real engineering organisation running a real migration on a production codebase and reporting the result.

Fable 5 shares its underlying model with Claude Mythos 5 — Anthropic's most restricted model, available only to vetted partners through Project Glasswing due to its cybersecurity capabilities. The difference between the two is not intelligence. It is guardrails. Fable 5 is Mythos made safe for general release — frontier capability wrapped in classifiers that block the highest-risk output categories and fall back to Claude Opus 4.8 when triggered. In practice, that fallback occurred in fewer than 5% of sessions.

The Every newsletter ran a seven-person internal test across a week and called it the best coding model in the world. Genspark said it beat every other model in its evaluations. Hex reported a 90% score on its core analytics benchmark of complex, long-running analytical tasks. The consensus among developers who got three days with it was emphatic: this was a generational step change, not an incremental improvement.

The timeline — three days from launch to blackout

What triggered the shutdown

The official reason was a jailbreak. Amazon researchers had found a sequence of prompts that could get Fable 5 to produce information potentially useful in cyberattacks. Amazon CEO Andy Jassy reportedly communicated this to the White House — including Treasury Secretary Scott Bessent — and the export control directive followed.

Anthropic's response was pointed. The company said it had reviewed what it believed was the basis of the directive and concluded that the capability demonstrated was widely available from other models and used every day by cyber defenders. In other words: the jailbreak did not unlock anything uniquely dangerous. It revealed capabilities that already exist across the industry.

The government never provided specific details of its concern beyond citing "national security authorities." Anthropic complied — because it had to — while making its disagreement unambiguous.

Dario Amodei: the man who warned about the wolf, then released it

The timing of Amodei's policy essay — published one day after Fable 5 launched, calling for governments to be empowered to block unsafe AI models — was not lost on anyone. Critics were brutal.

Yann LeCun — Turing Award winner, co-founder of AMI Labs after departing Meta, and the most prominent public sceptic of large language model supremacy — posted on Facebook that Amodei's "ridiculous fear mongering" was directly responsible for what happened. "One reaps what one sows," he wrote. LeCun has consistently argued that the safety warnings from frontier AI CEOs are overstated, self-interested, and damaging to public discourse. He believes the breakthroughs in AI will come not from scaling LLMs but from world models and architectures that understand physical reality — and that the current race is building on foundations that will plateau.

Adam Thierer of the R Street Institute acknowledged on X that Anthropic had "relentlessly raised the regulatory temperature in Washington" — but still argued the government's action represented "a significant escalation in the politicisation of AI and centralisation of control over advanced computation."

Alex Stamos, chief product officer at Corridor and former security chief at Facebook, penned an open letter co-signed by more than 150 executives calling for the directive to be lifted. His central argument: the rules governing AI model safety need to be written down and transparent. "There's nothing that Anthropic or anybody else can look at to say what can and can't I do," he told CNBC.

Stratechery's Ben Thompson offered the defence Anthropic would not say itself: "There's reason for cynicism, given Anthropic's history, but the part of the 'Boy Cries Wolf' myth everyone forgets is that the wolf did come in the end."

The voices — what the key figures actually said

The China problem

While Fable 5 was dark, something uncomfortable happened. Chinese open-weight models moved into the gap.

Z.ai's GLM-5.2 arrived right after Anthropic disabled access. As an open-weight model — meaning its trained weights are publicly downloadable and self-hostable — it cannot be switched off by a government letter. Benchmarks placed it close behind leading closed-source models. Reuters reported it surprised global users.

DeepSeek V4 advanced its context window to 1 million tokens as standard, adapted for Huawei Ascend chips, reducing dependence on foreign AI hardware.

The irony was pointed. The US government's action, taken in the name of national security, handed 19 days of uncontested access to Chinese AI providers. Enterprise developers, mid-task, facing broken pipelines, went looking for alternatives. Some found Chinese open-weight models. Some of those developers did not come back.

The deeper tension — and the Department of War

The Fable shutdown was not the first confrontation between Anthropic and the Trump administration. In March 2026, the Pentagon labelled Anthropic a "supply chain risk" — a designation historically reserved for US adversaries, never before applied to an American company — after Anthropic refused to provide Claude for what it considered ethically unacceptable military applications.

The Department of War had demanded Claude be available for all lawful uses, including mass domestic surveillance and fully autonomous weapons. Anthropic refused. A federal judge issued a preliminary injunction finding the designation likely violated constitutional protections. Litigation was still ongoing when the export control directive arrived.

Dario Amodei's statement on the Department of War dispute drew a line that few CEOs of $1 trillion companies would draw in public: "We cannot in good conscience accede to their request."

At the same time, Anthropic is not a neutral actor in Washington. It signed a $200 million Department of Defense contract in 2025. It was the first frontier AI company to deploy models in classified US government networks, the first at the National Laboratories, and the first to provide custom models for national security customers. Claude is extensively used across intelligence analysis, operational planning, and cyber operations.

The picture that emerges is complicated. Anthropic is simultaneously the company most trusted by the US government with classified AI access, the company most publicly committed to safety as a constraint on that access, and the company most willing to sue the government when it disagrees with how that access is demanded.

What this means for anyone building on AI

The Fable affair revealed something that developers and enterprises had not fully internalised: AI models are infrastructure, and infrastructure can be taken away by a letter.

Not by a bug. Not by a deprecation schedule with six months notice. By a government official, at 5:21 PM on a Friday, with 90 minutes to comply.

VentureBeat reported that enterprise technical leaders are already moving toward model-agnostic fallback architectures — proxy layers that can dynamically reroute production pipelines from proprietary APIs to locally hosted open-weight alternatives when access is disrupted. The lesson from June 2026 is that single-provider dependency on a frontier closed model is now a recognised risk category, not just a theoretical one.

For smaller operators — developers, agencies, startups — the implications are more immediate. If you have built a product that depends on a specific model's capabilities, you need a fallback. Not because the model will be bad, but because it might not be there.

Where Claude stands now

Fable 5 is back. It is available on Claude Platform, Claude.ai, Claude Code, and Claude Cowork. For Pro, Max, Team, and select Enterprise plans, it is included for up to 50% of weekly usage limits through July 7, after which it requires usage credits.

The new safeguard that Anthropic implemented during the shutdown blocks the specific jailbreak pattern that triggered the directive in 99% of cases. In the remaining 1%, Fable's output only includes previously discovered or already patched security flaws. Commerce's Center for AI Standards and Innovation tested both the original and new safeguards independently.

Anthropic's revenue run rate stood at $47 billion at the time of the shutdown, up from roughly $10 billion a year prior. The company recently closed a funding round at a $965 billion valuation — higher than OpenAI's $852 billion at the same time. Both companies have now confidentially filed IPO prospectuses with the SEC.

The question that does not yet have an answer is whether the regulatory framework will be resolved before the next Fable-class model arrives. The Trump administration faces an August 2026 deadline to create standardised benchmarks for evaluating security risks of new AI models. Whether that produces a coherent, written, transparent process — or another case-by-case intervention at 5:21 PM on a Friday — is the most important open question in AI right now.

LeCun may be right about the long arc. He may also be wrong about this specific moment. What is not in dispute is that a 19-day global blackout of the world's most capable public AI model, triggered by a government letter, is not a routine event in the history of technology. It is the first of its kind.

It will not be the last.