In the span of just three days, one of the most powerful AI models ever released to the public went live, drew global attention — and then went dark on government orders. The story of Claude Fable 5 is unlike anything the AI industry has seen before, and it raises questions that will define how humanity navigates the next chapter of artificial intelligence.
On June 9, 2026, Anthropic launched two models simultaneously: Claude Fable 5 and Claude Mythos 5. Fable 5 was described as a Mythos-class model — a tier above Anthropic's previous frontier Opus models — made safe enough for general public use. Mythos 5, the same underlying model but with safety classifiers removed in key areas, was restricted to a small group of vetted cybersecurity organizations through Anthropic's Project Glasswing.
Fable 5 was immediately declared the most capable AI model ever made available to the general public. It set state-of-the-art benchmarks across software engineering, knowledge work, vision, scientific reasoning, and long-horizon autonomous tasks. One real-world early test said it all: Stripe reported that Fable 5 compressed months of engineering work into days, performing a codebase-wide migration across 50 million lines of Ruby code in a single day — a task that would have taken a full team over two months by hand.
The model was not just faster. It was fundamentally different in kind from anything before it.
Three days after launch, on June 12, 2026 at 5:21 PM ET, the U.S. government issued an export control directive to Anthropic. The order came from the Commerce Department, signed by Secretary Howard Lutnick, citing national security authorities. It demanded the suspension of all access to Fable 5 and Mythos 5 by any foreign national — whether inside or outside the United States, including Anthropic's own foreign national employees.
The problem: Anthropic has no way to filter users by nationality in real time. So the company faced a stark choice — comply selectively, or shut both models down entirely. It chose full compliance and pulled both models offline for all customers worldwide.
This was the first time in history that a leading AI company publicly took a deployed, commercially released model offline due to direct federal government intervention.
Anthropic's own statement made its frustration clear. The company said its understanding is that the real trigger was not the export control logic itself, but a claimed jailbreak of Fable 5 — specifically, a "narrow, non-universal jailbreak" that allowed users to prompt the model into analyzing a codebase and identifying software vulnerabilities. Anthropic pushed back firmly, noting that this same level of capability is already widely available in other public models, including OpenAI's GPT-5.5, and is used routinely by cybersecurity professionals for legitimate defensive purposes.
Anthropic also pointed out that Fable 5's most critical safeguards run through independent classifier systems that operate separately from the model itself — meaning that even if someone bypassed a surface-level refusal, the deepest protections against the most dangerous outputs remained active. Despite this, the government order stood, and Fable 5 went dark.
To understand the stakes, you have to understand what Fable 5 was actually capable of — and what its unrestricted sibling, Mythos 5, demonstrated in controlled settings.
In cybersecurity, Mythos-class models reached a genuinely new threshold. Testing showed these models could discover vulnerabilities in every major operating system and web browser Anthropic tested. They scored 78% on ExploitBench, which measures real-world vulnerability discovery and exploitation, and 88% on Terminal-Bench for coding. Not just flagging issues — actually completing multi-step cyberattacks autonomously: reconnaissance, exploit development, lateral movement, defense evasion. This is what the cybersecurity world calls "agentic hacking," and no prior public AI model had demonstrated it at this level.
Without Fable 5's classifiers, a motivated attacker would not need a team of security researchers. They would just need an API key.
In biology and chemistry, the risks were equally sobering. Anthropic tested Mythos 5's ability to design adeno-associated viruses (AAVs) — the building blocks of gene therapy. Without being trained on the task, the model outperformed dedicated protein language models. That is a remarkable scientific achievement. It is also a dual-use capability: the same reasoning that can help design a life-saving gene therapy can, in the wrong hands, assist in designing something far more dangerous. Anthropic extended biology safeguards on Fable 5 precisely because the line between therapeutic and harmful applications in this domain is extremely thin.
In intelligence and influence, a model that can conduct novel genomics research autonomously for a week, outperform published academic models while being 100 times smaller, and generate novel scientific hypotheses confirmed by independent labs — that model is not just a tool. It is something closer to an autonomous scientific agent. Deployed without restrictions, the implications for disinformation, synthetic biology, and economic espionage would be profound.
The irony Anthropic itself acknowledged: the very safety warnings the company used to frame Mythos — that it was too dangerous to release broadly — attracted exactly the government scrutiny that ultimately pulled Fable 5 offline too.
Set aside the risks for a moment, because the potential benefits were equally extraordinary.
In its brief three days online, the early signals were already remarkable. In drug design, Mythos 5 internally accelerated aspects of the pharmaceutical development pipeline by roughly ten times. Scientists found it could autonomously execute the full workflow a human researcher performs — selecting binding sites, running protein design tools, recovering from failures — and yielded strong drug candidates for nine out of fourteen protein targets. Targets included immune checkpoints, neurodegeneration, and muscle disease.
In genomics, the model conducted novel autonomous research over more than a week, assembling single-cell data across millions of cells from 138 animal species and building a custom machine learning model that outperformed a recently published journal paper in Science.
In software engineering, the benchmark performance was the highest ever recorded on multiple frontier evaluations. The model could rebuild a web app from a screenshot alone, beat complex games using only vision, and handle long-horizon agentic tasks that previous models failed on.
The human impact at scale is difficult to overstate. Models of this caliber, applied to healthcare, could compress years of drug discovery. Applied to climate science, they could model interventions at a level of complexity no human team can sustain. Applied to education, they represent a genuinely personalized, senior-expert-level tutor available to anyone with internet access. Applied to legal and financial work, they can perform at levels that currently require entire professional teams.
Fable 5 was not a marginal improvement. It represented a step-change in what AI can do — and the breadth of problems it can address.
What makes the Fable 5 story so difficult is that both sides of the argument are right.
Anthropic is correct that the specific jailbreak cited by the government was narrow, non-universal, and represents a capability already present in other public models. The company is correct that pulling a model used by hundreds of millions of people because of a limited vulnerability sets a standard that would effectively freeze all frontier AI development. And it is correct that its classifier architecture provides meaningful independent protection even when a surface refusal is bypassed.
The government is also correct that Mythos-class models represent a qualitatively new threat surface. The ability to autonomously discover and exploit vulnerabilities across every major operating system, or to assist in the design of complex biological agents, is not the same as writing code or summarizing documents. These capabilities, in the hands of a sufficiently motivated nation-state actor or criminal organization, carry consequences that could affect millions of people. No prior commercial AI model has operated at this level. Treating it like all prior models would be a category error.
The real question — one this shutdown has forced into the open — is who should make these decisions, on what timeline, and with what transparency. The government acted unilaterally and in hours. Anthropic had no appeal process. The model was gone before most of its users even knew what had happened.
This is not a sustainable governance model for a technology of this consequence.
Anthropic has said it is working to restore access as quickly as possible. The company disagrees with the government's assessment but has complied with the directive. Fable 5 is described as "temporarily unavailable" — not permanently withdrawn.
But the episode has surfaced questions that the AI industry, regulators, and society will need to answer in the months and years ahead:
Claude Fable 5 existed publicly for seventy-two hours. In that time, it set records across nearly every benchmark the industry tracks, compressed months of engineering into days, and pointed toward a future where AI models autonomously advance human science. It also demonstrated — clearly — that the same power that can accelerate human flourishing can, without careful controls, be turned against it.
The shutdown of Fable 5 is not the end of this story. It is, in a meaningful sense, the beginning.
Sources: Anthropic official blog post on Claude Fable 5 and Mythos 5; TechCrunch; NBC News; Decrypt.
