Anthropic has formally launched Claude Fable 5, the first model built on its Mythos-class technology to be made broadly available to business customers and premium subscribers. The release on June 9 marks a shift for a model family the company had previously kept behind tight restrictions.
Months earlier, Anthropic had limited a Mythos-level model to a small group of organizations, citing concerns that its advanced capabilities could be misused. The company now says it has developed new safeguards that can automatically block dangerous requests in cybersecurity, biological research and other sensitive fields.
Performance gains over Opus 4.8
The main selling point of Fable 5 is improved capability. According to Anthropic, the model outperforms Claude Opus 4.8 by more than 10 percent on some benchmarks covering software engineering, research, analysis and knowledge work. In practical terms, the company says, that translates into solving more complex problems, writing better code and conducting deeper research.
A fallback approach to safety
Fable 5 takes an unusual approach to safety. When a user asks about cyberattacks, hazardous biological information or other sensitive topics, the system automatically switches from Fable 5 to Opus 4.8 in order to provide a more limited, safer response.
Anthropic has also introduced Claude Mythos 5, which is based on the same architecture but with several of those restrictions relaxed. This version is not available to the general public. It is being provided only to approved organizations under Project Glasswing, the company’s program for giving vetted partners access to its most capable models for security work.
Glasswing expands to about 200 organizations
According to reports, Anthropic recently added roughly 150 organizations to the Glasswing program, bringing the number with Mythos access to about 200. Those organizations use the model for cybersecurity research, threat detection and testing advanced defensive systems.
The launch signals that the AI industry is entering a phase in which the most powerful models are released to the public with layered safeguards rather than kept entirely private. Whether that strategy makes AI more useful without creating new risks is a question the coming months will test.
