On Tuesday, June 9, Anthropic introduced Fable 5 as the most capable artificial intelligence model it had ever made widely available. By Friday night, it was gone.
The company abruptly disabled Fable 5 and the more powerful Mythos 5 after the U.S. government ordered it to prevent any foreign national from accessing either model. The restriction extended beyond people in foreign countries to non-U.S. citizens working or living in the United States — including Anthropic's own employees who happen not to be U.S. citizens.
Anthropic said it could not practically comply without shutting down the models for everyone.
This extraordinary sequence of events left customers without access to a model that many had begun using for sophisticated coding, research, and business tasks. It also opened a larger argument about national security, government control of frontier AI, and how much trust businesses should place in cloud-based tools they do not ultimately control.
Here is what was known as of Monday morning, June 15.
A powerful model with unusual cybersecurity abilities
Fable 5 and Mythos 5 use the same underlying Anthropic model, but they were released for different audiences.
Fable included safeguards designed to prevent users from applying its advanced capabilities to potentially dangerous cybersecurity work. When certain requests triggered those protections, Anthropic said the task would be redirected to its less powerful Claude Opus 4.8 model.
Mythos, which Anthropic described as having the strongest cybersecurity capabilities of any model in the world, was available only to a limited group of approved organizations. Its earlier version had been used by cybersecurity researchers and government partners to identify vulnerabilities in widely used software.
Anthropic said Fable was particularly strong in software engineering, scientific research, analytics, and other complex assignments. The company priced it at $10 per million input tokens and $50 per million output tokens, making it considerably more expensive than its other generally available Claude models.
The company also acknowledged the risks. Anthropic said the model's cybersecurity capabilities, without safeguards, could potentially be misused to cause serious damage. Those safeguards quickly became the center of the dispute.
Amazon report reportedly triggered government action
According to reporting by Axios, later matched by Reuters and The Wall Street Journal, Amazon contacted federal officials Thursday night after researchers found a way to bypass restrictions and access portions of Mythos through Fable. Amazon is both Anthropic's primary cloud partner and one of its largest investors.
Amazon's dual role made the move an unusual one, raising the question of why a major backer would help set in motion a shutdown that abruptly cut off a flagship product from a company it had invested in. In a statement, Amazon said it is not unusual for governments to seek its guidance on potential security risks as a major cloud provider, and that it does not disclose the details of those discussions.
At least five other companies also contacted administration officials with concerns, Axios reported. Federal officials spent hours speaking with Anthropic on Friday before the government imposed export controls covering both models.
Anthropic reportedly received a call around 1 p.m. Eastern Time saying it had 90 minutes to disable Fable and Mythos because of a national security threat. The company said it pressed officials to explain the specific threat so it could address it, but the government held firm. A formal export-control directive from the Commerce Department arrived at 5:21 p.m., and users lost access to Fable around 10 p.m.
The two sides appear to disagree sharply about what Amazon's researchers demonstrated.
Anthropic reportedly argued that the technique was relatively simple, could be used with other AI models, and did not reveal a unique failure of Fable's safeguards. Katie Moussouris, CEO of cybersecurity firm Luta Security, told Axios that the government's response appeared disproportionate to the research and that the questions involved were similar to those legitimate cyberdefenders would ask.
An administration official countered that Anthropic had failed to treat the report seriously enough and should have paused access while investigating it.
Over the weekend, White House AI adviser David Sacks offered the administration's account on X, saying a partner trusted by both Anthropic and the government had found a jailbreak in Fable 5's guardrails and that officials asked Amodei to fix it or pull the model. When he declined, Sacks said, the government issued the export control reluctantly, and it hopes Anthropic will remediate the issue so Fable can return to general release.
The government has not publicly released the technical findings behind its decision. That makes it difficult to determine whether the shutdown addressed an exceptional vulnerability or established a far broader precedent for government oversight of powerful AI models.
Anthropic technical employees were traveling to Washington on Sunday to meet with White House officials, Axios and Reuters reported. Sources on both sides said they wanted to resolve the disagreement, leaving open the possibility that access could be restored relatively quickly.
A three-day warning for businesses
Christopher Penn, co-founder and chief data scientist at Trust Insights, tested Fable during its brief public availability.
Writing in his Almost Timely newsletter early Sunday morning, Penn called the model excellent, although "very, very expensive." He was particularly impressed by its ability to serve as a master reviewer — delegating lighter assignments to smaller models while examining a larger project for bugs and problems that earlier quality-control checks had missed.
Penn also questioned whether most users needed Fable's power for everyday assignments. With thoughtful requirements, detailed specifications, a work plan, and adequate time to review the results, he argued, less expensive models could accomplish many of the same tasks.
But his larger takeaway had little to do with model performance. "The events of this week proved that any government can unilaterally cut you off from AI services," Penn wrote.
For organizations whose operations increasingly depend on generative AI, he argued, private or locally operated AI should become part of a business-continuity plan. A company relying solely upon a cloud provider remains exposed to government intervention, provider policies, price changes, service outages, and other decisions outside its control.
Penn strongly favors running local AI models on company-owned hardware or privately managed infrastructure. Of course, that will not be practical for every organization, particularly when the most advanced models require expensive computing equipment and specialized technical knowledge.
The broader warning applies more widely: Companies should know which operations would be disrupted if a favored AI model disappeared tomorrow. They should also consider whether important workflows can move to another provider, continue with a less capable model, revert to conventional software, or operate locally.
Penn's concern is already echoing beyond business. The shutdown has prompted worry among UK and European leaders and has intensified calls for "sovereign AI" — the idea that countries should control the models and infrastructure their critical systems depend on, rather than rely on a provider that another government can switch off.
Until last week, the sudden disappearance of a leading commercial model would have seemed far-fetched. Fable 5 lasted three days – long enough for companies to start building on it, but not long enough to count on it.
Editor’s note: This is a developing story. Colorado AI News will update this article as additional, material information becomes available.
