Anthropic relaunches Claude Fable 5 after controls lifted

Anthropic will offer Claude Fable 5 again to users worldwide starting July 1, after the United States lifted the export controls that had blocked access to the model for 18 days.

The company suspended Fable 5 and Mythos 5 for all users on June 12 because it could not verify each person’s nationality in real time. The measure also affected foreign nationals using the models inside the United States.

What changes for users

Fable 5 will return to Claude Platform, Claude.ai, Claude Code and Claude Cowork. Anthropic also plans to restore access on AWS, Google Cloud and Microsoft Foundry as soon as possible.

For Pro, Max, Team and some Enterprise plans, the model will be included for up to 50% of the weekly usage limit through July 7. After that, access will work through usage credits. Standard Enterprise accounts will not have an included allowance, so users will need to activate those credits to keep using the model.

Access to Mythos 5, which has fewer safety barriers and is designed for defensive cybersecurity, has been restored for some US organizations with official approval dated June 26. Anthropic is still negotiating to extend that access to domestic and international partners in the Glasswing program.

The problem that caused the block

The US government’s decision came after a report from Amazon researchers. The team found a way to induce Fable 5 to identify software vulnerabilities and, in one case, generate code showing how to exploit them.

Anthropic says its own tests changed the initial interpretation of the incident. Less capable models, including Claude Opus 4.8, GPT-5.5 and Kimi K2.7, were also able to find the same vulnerabilities. In addition, every model evaluated managed to produce the specific exploit demonstration.

According to the company, this means the method did not reveal unique offensive capabilities in Mythos 5. The behavior fell into a borderline area of Fable 5’s safeguards and involved a relatively routine defensive cybersecurity task.

Even so, Anthropic trained a new safety classifier, an automated system that detects potentially dangerous requests or responses. The company says it blocks the technique described in the report in more than 99% of cases. When the block is triggered, the request will be sent to Opus 4.8.

The cost will be more false positives: some legitimate programming and debugging requests could be rejected by mistake. Anthropic says it will keep adjusting the system to reduce those interruptions.

A common framework for measuring jailbreaks

Anthropic is also proposing that the industry adopt a shared way to measure the severity of a jailbreak, meaning a technique for bypassing a model’s safety barriers. The proposal, developed with Amazon, Microsoft, Google and other partners, would assess four areas:

• Capability gain: how much more the model can do compared with existing tools.
• Breadth: how many different offensive tasks the same method makes possible.
• Ease of malicious use: how much effort and knowledge the attacker needs.
• Ease of discovery: whether the technique requires specialized knowledge or is already publicly available.

The goal is to distinguish a limited flaw from a technique that unlocks many harmful capabilities with a single instruction. Anthropic is also preparing a HackerOne program to receive reports from researchers about potential Fable 5 jailbreaks.

More oversight from the US government

As part of the agreement, Anthropic will give certain agencies early access to models relevant to national security. Those teams will be able to test both the capabilities and the safeguards before the general release.

The company also promises to quickly share information about significant jailbreaks, collaborate on joint investigations and work with the government and other companies on a voluntary safety standard for advanced models.

For you, the immediate effect is simple: Fable 5 is coming back, but with stricter controls and an access system that will depend increasingly on your plan and usage credits. What matters now is whether this agreement becomes a stable process for releasing powerful models, instead of every incident ending in sudden blocks and different decisions for each company.