Anthropic brings Claude Fable 5 back online

Anthropic’s flagship Claude Fable 5 is back online, two weeks after a sudden regulatory crackdown forced the company to pull the plug.

The U.S. Department of Commerce lifted export controls on both Claude Fable 5 and Claude Mythos 5, paving the way for a global redeployment on July 1 across Claude.ai, Claude Code, Claude Cowork, and developer APIs.

The Amazon bypass that triggered Washington

The trouble started on June 12, just three days after Anthropic launched both models. While Fable 5 was built for the public with extensive guardrails, Mythos 5 was an unconstrained variant reserved strictly for trusted partners in Project Glasswing doing defensive cybersecurity.

Amazon researchers discovered a safeguard bypass in Fable 5. By carefully prompting the model, they coaxed it into identifying several software vulnerabilities and producing an exploit demonstration for one of them. Once alerted, the U.S. government issued an export control directive requiring Anthropic to block access to all foreign nationals immediately. Unable to reliably verify citizenship in real time, Anthropic simply shut off access worldwide.

Anthropic quickly red-teamed competing models to put the findings in context. The results showed that identifying the vulnerabilities wasn't unique to Fable 5—Claude Opus 4.8, GPT-5.5, and Moonshot AI's Kimi K2.7 all produced identical findings. Even more telling, every single model Anthropic tested, including older systems like Claude Haiku 4.5, Sonnet 4.6, and GPT-5.4, was able to generate the exact same exploit demonstration.

Over-defensive by design

To satisfy regulators at the Commerce Department's Center for AI Standards and Innovation (CAISI), Anthropic trained a new safety classifier that shuts down the specific bypass technique in over 99% of cases.

When a prompt trips the new security filters, Fable 5 refuses the request and silently routes the user back to Opus 4.8. Anthropic candidly acknowledged the catch: in broadening its "safety margin," the new filter will inevitably generate more false positives during everyday coding and debugging workflows.

Access rules are also shifting. Through July 7, Pro, Max, Team, and select Enterprise subscribers can allocate up to 50% of their weekly allowance to Fable 5. After that date, accessing the model will require separate usage credits. Mythos 5 access remains tightly controlled, restored only to select U.S. organizations.

Why it matters

This two-week saga marks the first time U.S. export controls were weaponized against a domestic frontier model's prompt safeguards—and it sets a chaotic precedent.

In our view, the episode highlights the absurd grey zone modern frontier models live in. The line between routine penetration testing and malicious exploit generation is razor-thin. When Amazon's bypass triggered an immediate national security reflex, it forced Anthropic, Google, Microsoft, and Amazon to quickly band together to establish a shared industry standard for classifying jailbreak severity. Without objective metrics, every frontier release risks being held hostage by subjective panic.

For developers, the redeployment is a mixed bag. You get frontier intelligence back, but you will almost certainly suffer from heavy-handed guardrails. If you rely on Claude for intensive code auditing or reverse engineering, expect constant pushback from the safety filters and unwanted reroutes to Opus 4.8 until Anthropic dials down its false-positive rate.