Fable 5 was meant to be the most powerful consumer model Anthropic had ever launched. Three days later, it is cut off. On 12 June 2026 at 23:21, Paris time, Anthropic receives a US directive: block Fable 5 and Mythos 5 for any foreign national. Unable to sort users in real time, the company suspends both models for everyone.
This is not an outage. It is a precedent: a commercial AI model, already launched, is treated as a strategic technology.
The short timeline
9 June 2026, 19:00 — Public launch.
The Verge publishes the launch of Fable 5, the first widely accessible Mythos-class model.
10 June 2026 — Jailbreak claimed.
CybersecurityNews reports that Pliny the Liberator claims a public jailbreak of Fable 5.
11 June 2026, 09:18 — Guardrails corrected.
Business Insider reports that Anthropic corrects invisible guardrails after criticism.
12 June 2026, 23:21 — US directive.
Anthropic says it received the US export control directive at 17:21 ET, that is 23:21 in France.
13 June 2026, 03:59 — API cut-off observed.
Simon Willison observes the Fable 5 API failing: the model is no longer available.
13 June 2026 — Cut-off confirmed, national security framing.
Anthropic confirms the suspension of Fable 5 and Mythos 5. Reuters, AP and Axios place the cut-off within a US directive tied to national security and export control.
Why Fable 5 was cut off
According to Anthropic, the US government invokes national security and a possible jailbreak method. The directive targets any foreign national, including inside the United States and including some of Anthropic’s own foreign employees.
The problem: Anthropic says it cannot distinguish those users in real time. Practical result: Fable 5 and Mythos 5 are cut off for all customers. The other Claude models, including Opus 4.8, are not affected.
The case is critical because it does not target a chip or a server, but the model itself. Axios explains that the United States is here imposing access control on two Anthropic models considered sensitive enough to fall under export control logic. That is a shift: software becomes a strategic capability.
To understand what Fable 5 represented before it was cut off, see our earlier article on Claude Fable 5 and its link to Mythos.
The incidents after release
First incident: the invisible guardrails. On 11 June, Business Insider reports that Anthropic is reversing a contested choice: some requests tied to AI development were degraded or redirected with no clear explanation. Anthropic then acknowledges it made the wrong trade-off and promises to make refusals or switches visible.
Second incident: refusals that were too broad. The Verge reports that Fable 5 could refuse or redirect very basic biology questions to Opus 4.8. The problem is not that Fable 5 did not know how to answer. The problem is that Anthropic was preventing it from answering, out of caution.
Third incident: the claimed jailbreak. On 11 June, CybersecurityNews reports that a red-teamer, Pliny the Liberator, claims to have bypassed Fable 5’s guardrails using framing, Unicode, decomposition and recomposition techniques. The article also mentions a leak of the system prompt, estimated at around 120,000 characters, and outputs tied to technical exploitation of the buffer overflow type on Linux x86.
Anthropic nonetheless disputes the seriousness of the signal used by the government. In its statement, the company says it saw only a narrow, non-universal jailbreak, concerning the reading of one specific codebase and the fixing of simple or already known flaws. It adds that no harmful result was demonstrated.
That is exactly where the case becomes dangerous: nobody really disputes that the risk exists. The disagreement is about the acceptable level of risk.
The Anthropic paradox
Anthropic spent months explaining that Mythos was too powerful to be released without strict control. On 9 June, it launches Fable 5 as the “secured” consumer version. On 12 June, the US government answers in practice: if it is that sensitive, we cut it off.
The paradox is brutal: Anthropic’s safety rhetoric turns against it.
The company wanted to prove it could deploy a highly advanced model with guardrails. The state has just shown it could take that decision out of the company’s hands.
This subject goes beyond Fable 5. It concerns every model that works for longer, reasons more deeply, reads code, uses tools and becomes more autonomous. We already saw it with Claude Opus 4.8 and long tasks, then with AI agents able to act on real systems.
What this changes for users
Fable 5 shows a simple truth: in cloud AI, you do not own the model. You rent access. That access can disappear for a commercial, technical, regulatory or geopolitical reason.
For an individual, that is frustrating. For a company, it is an architecture risk. If a workflow, a product or a business tool depends on a single model, one directive can break everything within hours.
The signal is clear: you need backup models, fallbacks, and a strategy less dependent on a single supplier. Le Recul’s AI ranking exists precisely to compare models by use case, not only to follow the latest fashionable model.
Our verdict
The suspension of Fable 5 is not only a jailbreak story. It is the first big collision between three forces: the commercial race between AI labs, the security fears of states, and users’ dependence on cloud models.
Anthropic says the measure is disproportionate. The US government answers with facts: it has the power to cut.
The most worrying part is not that Fable 5 was suspended. The most worrying part is that nobody yet knows where to draw the line: at what level does an AI stop being a product and become a strategic capability?
What to take away
Fable 5 launches on 9 June 2026.
A public jailbreak is claimed around 10 June.
Anthropic corrects its invisible guardrails on 11 June.
The US directive arrives on 12 June at 23:21, Paris time.
Fable 5 becomes unavailable on the API side around 13 June at 03:59, Paris time, according to Simon Willison’s tests.
The other Claude models remain accessible.
The case sets a precedent: an AI model can be cut off not because it breaks down, but because a state judges it too sensitive.
The figure to remember
3 days.
That is the time between Fable 5’s public launch and its suspension after a US directive.
In frontier AI, three days can be enough to go from “revolutionary new model” to “strategic capability under state control”.
The line to remember
Anthropic wanted to prove that a very powerful model could be made safe by guardrails. The US government has just pointed out that beyond a certain level of power, private guardrails are not always enough.