Anthropic Relaunches Claude Fable 5 on July 1 After Lifting US Export Controls, Adds New Cybersecurity Classifier

Anthropic redistribution The legend of Claude 5its the most efficient model available in general. On June 30, it announced that US export controls had increased. The controls were covered The legend of Claude 5 again Claude Mythos 5. Mythos 5 returned to users worldwide on Wednesday, July 1. Access to Mythos 5 is restored to a set of US organizations.

The models were pulled on June 12. The US government order allowed only non-immigrants. Anthropic could not verify nationality in real time. So set up both models for everyone.

This article explains what caused the blockage. It includes new protections and a proposed jailbreak framework. It also shows how the Fable 5 compares to rivals like the GLM-5.2.

Fast facts

• Model: Claude Fable 5 (Mythos-class model made safe for general use)
• Event: It was reinstated on 1 July 2026 after export controls were lifted
• Reason for suspension: Amazon’s report on security breaches
• Fix: A new security feature that blocks the reported process
• Price: $10 million input tokens, $50 million output tokens
• Where: Claude Platform, Claude.ai, Claude Code, Claude Cowork

What happened: timeline

Anthropic launched Fable 5 and Mythos 5 on June 9. Both have the same basic model. 5 legendary ships with strong defenses for general use. Mythos 5 has some protections suggested by our cybersecurity partners.

On June 12, the US government implemented export controls. The order went into effect immediately. Suspended access instead of non-compliance risk.

The trigger was a report from Amazon researchers. They found a way to bypass the protections of Fable 5. The information enabled the model to identify a number of software vulnerabilities. In one case, it generated code that showed how to exploit a single vulnerability.

On June 26, the government authorized the return of Mythos 5 to some US entities. On June 30, the controls were completely removed.

Why Anthropic says the findings were not unique

Anthropic checked to see if that finding was different from Fable 5. It wasn’t.

Models can effectively identify the same vulnerability. That list includes the Claude Opus 4.8, GPT-5.5, and Kimi K2.7.

In one manipulation demonstration, every tested model reproduced. That set included Haiku 4.5, Sonnet 4.6, Opus 4.6, and Opus 4.7. It also covers Opus 4.8, GPT-5.4, GPT-5.5, and Kimi K2.7.

The Anthropic team says this approach did not produce Mythos-level cyber skills. Call this case the Fable 5 security boundary. The blocked behavior only involves normal cybersecurity work.

How the new separator works

Anthropic has moved to fill the gap. Trained an advanced safety class for reported behavior.

A separator blocks a particular method in more than 99% of cases. Blocked requests are not rejected outright. They have been ported to Claude Opus 4.8 instead. Users are notified when this backlog occurs.

Commerce Department CASI researchers tested old and new defenses. They agree that the defenses are incredibly strong. The tradeoff is more false positives during normal coding and debugging.

This shows the Anthropic structure of ‘defense in depth’. Classifiers are small AI systems that detect malicious cyber activities. A deliberate ‘safety margin’ also prevents certain humanitarian requests. The Legend 5 uses a much larger safety margin than previous models.

A proposed framework for jailbreak robustness

The episode revealed a gap. The industry does not have a shared standard for finding a ‘jailbreak,’ a method that bypasses the model’s protections.

Anthropic is writing one with Amazon, Microsoft, Google, and other Glasswing partners. The draft gets a jailbreak in four conditions:

1. Gaining strength – how far it takes from the tools available to the user.
2. Range of skill advantage – how many different annoying functions it opens.
3. Ease of use of weapons – how much human effort the attack still requires.
4. Availability – how easily one can find this program.

On the hardest class, Anthropic will apply the first reduction immediately. It also offers 24/7 monitoring of jailbreak channels.

Interactive scorer

Try this interactive embedded scorer to see how these four criteria fit together.

Use cases with examples

Legend 5 targets a long horizon, an agent’s career. Here’s where first-time developers can use it.

• Codebase Migration: Stripe reported an extensive codebase migration in one day. The work included a Ruby codebase of 50 million lines. Doing it by hand will take the team more than two months.
• Financial analysis: In Hebbia’s Finance Benchmark, Fable 5 scores high. It benefits from chart, table, and document visualization.
• Idea-to-code: Legend 5 can reconstruct the source code of a web application from screenshots alone.
• Agents work for a long time: File-based memory helps it stay focused on millions of tokens.

How does Fable 5 compare?

The suspension created an opening for competitors. Days after the suspension, Zhipu AI released the GLM-5.2 as open weights. Independent testers rated it as the most robust model available.

GLM-5.2 uses a Mixture-of-Experts design. It has about 750 billion parameters. Only about 40 billion worked for each token. In Semgrep’s IDOR benchmark, it scored 39% F1. That beats Claude Code by 32% over the same period.

The gap is narrowing in cost. In AA-Briefcase, Legend 5 averages $31 per transaction. The GLM-5.2 estimate is $2.40.

Access and quick API example

On Pro, Max, Team, and select Enterprise plans, Fable 5 is included until July 7. Includes up to 50% of weekly usage limits. After that, access moves to usage credits. Anthropic also powers Fable 5 on AWS, Google Cloud, and Microsoft Foundry.

The developers call the model the claude-fable-5 string:

from anthropic import Anthropic

# Reads your key from the ANTHROPIC_API_KEY environment variable
client = Anthropic()

message = client.messages.create(
    model="claude-fable-5",
    max_tokens=1024,
    messages=[
        {"role": "user", "content": "Refactor this module for readability."}
    ],
)

print(message.content)

If the separator fires, the answer comes from Opus 4.8. Your code path remains the same.

Important takeaways

• Legend 5 returns on July 1st after export controls are lifted.
• The new category prevents the reported passing of more than 99% of cases.
• Blocked requests route to Opus 4.8, not outright rejection.
• Anthropic suggests a four-criteria framework for jailbreak scoring.
• The GLM-5.2 has emerged as a cheap open-weight contender for pause time.

Check out Technical details. Also, feel free to follow us Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to Our newspaper. Wait! are you on telegram? now you can join us on telegram too.

Need to work with us on developing your GitHub Repo OR Hug Face Page OR Product Release OR Webinar etc.? Connect with us