Anthropic releases Claude Fable 5 to the public, its most capable model yet, with an automatic safety system that diverts high-risk queries

Public release of a restricted powerhouse

Anthropic on Tuesday (9 June) made Claude Fable 5 generally available, marking the first time a model from its advanced Mythos class has been released to the public. The Mythos line was unveiled in April but kept behind closed doors because of its ability to spot cyber vulnerabilities in critical infrastructure—including banking platforms and power grids—that elude human detection. The same week, Anthropic had published a document calling on major labs to consider a coordinated AI slowdown, describing the possibility of “slowing down or temporarily pausing” global progress as “probably a positive thing.”

Launching such a powerful model carries risks. Without protections, Fable 5’s abilities in fields like cybersecurity could be misused to cause serious harm.

How the safety filter works

Fable 5 incorporates what the company describes as a “double track” or “air traffic control” system. When a user’s query enters a predefined high-risk area—cybersecurity and biology are the two most guarded domains—the model does not answer directly. Instead, the request is handed off to Claude Opus 4.8, a less capable but safer model. Anthropic says this approach, an extension of its Constitutional AI framework, lets the model operate at full power on most tasks while capping output on sensitive subjects. The risk concern is twofold: automated exploitation of software vulnerabilities and misuse in the life sciences.

Performance leap over previous models

Benchmarks show a generational jump. On SWE-Bench Pro, which measures autonomous programming, Fable 5 scored 80.3% against 69.2% for the previous best public model, Claude Opus 4.8, with the gap widening on longer and more complex tasks. Analytics firm Hex reported that Fable 5 became the first model to surpass 90% on a benchmark for complex, long-duration analytical tasks—a ten-point improvement over the Opus generation. Anthropic says the model “exceeds the capabilities of any model we’ve ever made available,” pointing to its strength in writing and debugging code, answering complex research questions, and analyzing images.

Fable 5 exceeds the capabilities of any model we’ve ever made available.

Separate tracks: Fable for all, Mythos for partners

Parallel to the public release, Anthropic is offering an unrestricted version called Claude Mythos 5 to organizations already granted access to the Mythos class. These are primarily cybersecurity partners enrolled in Project Glasswing, an initiative launched in April 2026 to use advanced AI to protect software and critical infrastructure. In early June, membership grew to roughly 200 organizations across more than 15 countries. The US government, which had been in a legal dispute with the firm, also tested the model due to security concerns and has since established an agreement with the White House on testing powerful AI models.

Contradictions and concerns

The timing has drawn criticism. Days before the release, Anthropic’s public document urged an industry-wide slowdown, warning that frontier models could soon begin to improve themselves without human supervision. Releasing a model from the very Mythos class that sparked these fears has led some observers to question the company’s consistency. Earlier, when Project Glasswing was first announced, critics had accused Anthropic of exaggerating the cyber threat as a marketing tactic. The company maintains that Fable 5’s safeguards make it safe for mass use, while the unrestricted Mythos 5 remains strictly gated.