After an export ban and an AI jailbreak, Anthropic commits to a new Claude Fable 5 strategy

Anthropic has confirmed that Claude Fable 5 will become a permanent part of the Max and Team Premium plans starting July 20. It caps a turbulent debut marked by U.S. export restrictions, a temporary global takedown of the model, a security probe into a so‑called jailbreak, and multiple extensions of its free availability. The revamped plans aim to give users clarity while Anthropic works to rein in the steep costs of its most powerful AI model. The official announcement went live on the Claude X account and follows a series of updates expanding the company’s infrastructure and security.

The move is a pivotal moment for Anthropic—not only because Fable 5 now has a permanent home in the top-tier plans, but also because the past weeks exposed just how hard it has become to make frontier AI both safe and globally available.

Starting July 20, Anthropic will roll out a new subscription lineup for Claude Fable 5.

Claude Max and Team Premium subscribers will get standard access to Fable 5. Within these plans, the model can be used for up to 50% of the weekly usage limit.

For Claude Pro and Team Standard, access shifts. They’ll keep access to Fable 5, but only via usage credits. To ease the transition, existing users will receive a one-time $100 credit so they can continue using the model without immediate extra costs.

Anthropic says this is the final subscription structure after demand for Fable 5 far exceeded expectations.

The rollout of Claude Fable 5 was anything but typical.

Anthropic introduced Fable 5 in early June as one of its most advanced AI models to date. At the same time, it unveiled Mythos 5, an even more capable model made available only to a small circle of trusted partners within Project Glasswing for defensive cybersecurity uses.

Days later, everything changed.

On June 12, the U.S. government unexpectedly imposed export restrictions on both Claude Fable 5 and Claude Mythos 5, barring Anthropic from making the models available to foreign nationals.

Because the restrictions took effect immediately—and Anthropic said it had no reliable way to verify users’ nationality in real time—the company made an extraordinary move: it temporarily shut down access to both models worldwide, including for U.S. users. The decision sparked confusion across the AI community.

The export move wasn’t prompted by a data breach or cyberattack, but by a security investigation.

According to Anthropic, Amazon researchers discovered a method that could partially bypass Claude Fable 5’s safety measures. With a specific prompt, the model identified software vulnerabilities and, in one case, even produced example code showing how a flaw could theoretically be exploited.

That raised concerns within the U.S. government, which imposed temporary export limits until the situation could be fully evaluated.

Anthropic then worked closely with the government, Amazon, and other partners to investigate the findings.

A striking conclusion from the review: according to Anthropic, Fable 5 did not possess unique offensive cyber capabilities.

In internal tests, Anthropic ran the same prompts on multiple competing AI models. GPT‑5.5, Claude Opus 4.8, and Kimi K2.7 were among the models that identified the same vulnerabilities as Fable 5.

Anthropic also says nearly all tested models produced the same kind of demonstration code.

The company concluded the reported jailbreak did not expose any exclusive features of Mythos 5, but was a borderline case within Fable 5’s existing safety guardrails.

While downplaying the severity, Anthropic still tightened security.

It built a new AI classifier specifically trained to detect and automatically block the described jailbreak scenario.

Anthropic says this added layer now stops the attack in over 99% of cases.

If a request is flagged as potentially harmful, users are notified and the task is automatically handed off to Claude Opus 4.8, which enforces stricter safety constraints.

Anthropic admits the tighter security has a downside. Legitimate coding and debugging requests may be wrongly blocked more often, meaning developers sometimes need extra steps to get work done.

A key piece of the new strategy is defense in depth.

Instead of relying on a single safeguard, Anthropic layers multiple defenses that reinforce each other.

The model is trained to refuse dangerous prompts, while separate AI classifiers continually scan for attempts to bypass safeguards. Suspicious usage patterns are also reviewed after the fact, and new detection rules can be rolled out quickly when fresh jailbreak techniques appear.

According to Anthropic, no AI model is fully immune to jailbreaks. The goal isn’t zero incidents, but making abuse as difficult as possible.

The Fable 5 episode exposed another gap, Anthropic says.

There’s no common standard to assess how severe an AI jailbreak is.

Anthropic will therefore work with Amazon, Microsoft, Google, and others on a shared evaluation framework.

The framework will consider factors such as:

The endgame: help AI companies and governments quickly triage which vulnerabilities demand immediate fixes and which pose lower risk.

Anthropic also announced an expanded partnership with the U.S. government.

Future frontier models will reach selected agencies earlier for independent safety testing. The company will also share information more rapidly on new jailbreaks, emerging threats, and mitigations.

Anthropic is setting aside additional compute for joint AI safety research and aims to help shape a voluntary international safety standard for advanced AI models.

This week’s announcement shows Anthropic hasn’t just tightened security—it has also scaled the necessary infrastructure.

During the trial window, demand for Fable 5 far exceeded expectations. The company repeatedly extended free access while adding GPU capacity.

With safeguards reinforced and infrastructure largely scaled up, Anthropic says it’s time to fold Fable 5 permanently into the Max and Team Premium plans.

For Pro users, the company is opting for a usage-credits model. That lets Anthropic control costs for its most expensive AI model without cutting off access entirely.

Claude Fable 5 signals a shift in the AI industry.

A few years ago, the race was mostly about raw model performance. Today, three forces dominate: compute, safety, and regulation.

Temporary export limits, jailbreak research, and a new subscription model all underline that frontier AI is no longer just a tech product—it’s also a geopolitical and economic issue.

For users, the new subscription structure brings clarity. For the industry, it shows that powerful AI increasingly emerges from tight coordination between tech firms, cloud providers, and governments.

Anthropic’s approach previews how next-gen AI may roll out: phased, heavily secured, and under growing government oversight.