Things in AI keep developing at breakneck speed. Just as we got an exciting new announcement from Anthropic about Claude Fable 5 being made available to the public, the first Mythos-class model anyone could actually use, the whole excitement was shut down by another announcement just days later. On June 9, Anthropic celebrated the launch. By June 12, the US government had ordered it gone. What happened in those 72 hours reveals more about AI power, trust, and regulation than months of policy debates. Anthropic launched Claude Fable 5 across multiple platforms: the Claude API, AWS Bedrock, Google Vertex AI, Microsoft Foundry, and Claude itself. It was the first publicly available ‘Mythos-class’ model, the same line of technology that Anthropic had held back in April because of its power and potential danger. Fable 5 had the same core capabilities as Mythos 5 but with hard safety limits, making it safe for general use. The launch included limited-time free access for paid subscription plans. For Pro, Max, Team, and Enterprise customers, Fable 5 was included at no extra cost until 22 June 2026. After June 23, continued access would have required usage credits billed at API token rates. Anthropic said it was widely released and built for demanding reasoning and long-horizon agentic work. But on June 12, Anthropic received a directive from the US government ordering them to suspend all access to Fable 5 and Mythos 5. By the evening, both models were disabled for all users worldwide. As of June 14, they remain unavailable. Anthropic says it’s working to restore access as soon as possible, and users are being redirected to Claude Opus 4.8.
Why did the government act so quickly? According to reports, the directive was triggered by a demonstration in which Fable 5 read a codebase and fixed vulnerabilities. US officials interpreted this as the model bypassing safety controls, a ‘jailbreak‘ that could let attackers use the same capabilities to compromise systems, steal data, or cause harm. The model’s advanced capabilities in cybersecurity, biology, and chemistry create what Anthropic itself calls a real risk of misuse, since the same skills that help defenders could help attackers. Anthropic’s response has been different. The company says the government misunderstood the demonstration. The vulnerabilities were already known and minor. Other models, including GPT-5.5, can find the same vulnerabilities without any safety bypass. This wasn’t a jailbreak, according to Anthropic, but normal model behaviour that was misinterpreted. This disagreement opens up three possibilities, and we do not need to pick one to see what’s wrong: Possibility 1: Low AI understanding among officials. Government agents may not have sufficient technical depth to distinguish between ‘model doing what it is designed to do‘ and ‘model bypassing controls‘. If a code-fixing demonstration looks alarming to someone who doesn’t understand AI architecture, panic becomes the default response. This mirrors broader concerns about regulatory capacity. Can governments even evaluate what they’re trying to control? Possibility 2: Corporate spin from Anthropic. The company has reputational and financial stakes in keeping Fable 5 public. Calling it a misunderstanding is strategically convenient because it minimises their responsibility and pressures the government to reverse course. They may be downplaying real risks to regain public access. Possibility 3: Something in between. The government may have overreacted based on incomplete information. Anthropic may have oversimplified their explanation to gain public support. Both sides may be talking past each other because they use different frameworks, security versus innovation. Regardless of who’s right, this reveals a structural gap. There is no established process for the joint evaluation of high-capability models prior to release. The government sees something alarming and acts. The company says, ‘Wait, you misunderstood ‘. But by then, the model is already shut down. This is the reactive regulation problem in action.
The Fable 5 case is a textbook example of reactive regulation. Ideally, what we’d want is proactive regulation. Why does this matter? Reactive regulation is unpredictable. Companies can’t plan their releases. The public gets whiplash. Innovation suffers from uncertainty. Proactive regulation is more stable, with clearer rules of engagement and better trust between sectors. Here’s the contradiction. The USA has bragged about avoiding AI regulation to protect innovation. The current administration has been vocal about not wanting to drown AI innovation with regulation and has even blocked some state-level AI regulation attempts. But Fable 5 shows that when capability crosses into security concerns, the government doesn’t need legislation. It uses export controls. This is not proactive regulation; it is reactive power. The EU takes a different approach. The EU AI Act is proactive, even if flawed, because it establishes rules before capabilities emerge. The USA is now demonstrating security-based reactiveness, with no legislation and relying on export controls that shut down models after someone raises an alarm instead. The question is whether this is better or worse than the EU’s approach. Export controls are faster and more flexible than legislation, but they are also less predictable, less transparent, and more prone to misunderstanding.
This is where the question of humanism becomes unavoidable. What happens when AI capabilities outpace trust between companies and governments? How does this affect public access to transformative technology? Mythos 5 is the most powerful model Anthropic has ever shipped, with twice the capability of Opus 4.8. This is transformative technology. When Fable 5 was released, it was supposed to be ‘Mythos for wide audiences’, safe, guarded, and public. Researchers could use it. Small companies could build on it. General users could experience what the latest AI could do. The very etymology of Anthropic’s tiering tells the story. Mythos is the ancient, gatekept truth held by the elite; a Fable is a story shared with the common people. By grounding Fable 5, the state has effectively locked the public out of the highest tier of human reasoning tools. The tension is between AI as a public good, meaning open access, democratised knowledge, broad benefit, and AI as a security risk, meaning restricted, controlled, and kept from the public. When governments shut down powerful models reactively, who loses? Researchers who need the capability to study it. Small companies that built products around it. General users who finally got access to the most advanced AI. The question of humanism is simple and uncomfortable. Should humanity’s most powerful tools be accessible, or locked away when they scare someone? Fable 5’s suspension may be temporary. The misunderstanding may be resolved. Anthropic may restore access once they’ve clarified the situation with the government. But the change is real. Export controls may become the default AI policy tool in the USA. This is not regulation in the traditional sense. Export controls operate under different legal authority, namely national security. The administration can still say it is not regulating AI while using security powers to control it. What does this mean for future model releases? The global dimension matters too. Other countries may follow the US export control approach. If the USA uses security directives to shut down models, other governments may do the same for their own reasons. Fable 5 was supposed to bring Mythos to the public. Instead, it became a case study in how AI power is controlled when it scares governments. Fable 5 may return. Anthropic has said it is working to restore access and may extend free access if capacity improves. But the question remains: When AI capabilities cross into security concerns, what’s the right response? Reactive export controls? Proactive legislation? Or something else entirely. The answer will determine whether humanity benefits from transformative AI or remains locked in the dark.
Author: Slobodan KovrlijaWhat Happened: The timeline
Alleged jailbreak vs ‘misunderstanding’
AI regulation: Reactive vs Proactive
What happens when AI capabilities outpace trust between companies and governments?
The cost of governing by alarm
How is AI power controlled when it scares governments?
