Anthropic Unveils Fable 5, Keeps Powerful Mythos Capabilities Under Tight Control
By CU Today Staff —
SAN FRANCISCO—Anthropic has launched Claude Fable 5, a new frontier artificial intelligence model the company says delivers Mythos-class performance in software engineering, advanced reasoning and scientific research, while keeping its most powerful cyber and biomedical capabilities restricted to a limited group of vetted organizations, according to BankInfoSecurity.
The new model is broadly available through Anthropic's API and enterprise offerings and is priced at $10 per million input tokens and $50 per million output tokens. Anthropic told BankInfoSecurity that Fable 5 is designed for complex, long-running tasks that can span days or even weeks, with the ability to maintain context across millions of tokens and leverage persistent memory more effectively than previous Claude models.
According to BankInfoSecurity, Anthropic's more capable Mythos 5 model uses the same underlying technology as Fable 5 but has safeguards removed around cybersecurity and biomedical research. Access to Mythos 5 is currently limited to approximately 200 organizations that have been vetted through Anthropic's Project Glasswing program, which was created to provide advanced AI capabilities to trusted critical infrastructure and research organizations.
Anthropic cited early results showing significant gains in software engineering and business analysis. BankInfoSecurity reported that payments giant Stripe used Fable 5 to complete a migration across a 50-million-line Ruby code base in about one day, a project that previously would have required a team of engineers working for more than two months. Anthropic also said the model achieved leading scores on benchmarks measuring coding, financial reasoning, document analysis, chart interpretation and problem-solving.
At the same time, Anthropic is taking steps to prevent advanced AI cyber capabilities from being misused. BankInfoSecurity reported that Fable 5 automatically routes cybersecurity-related queries to the less capable Claude Opus 4.8 model, while the company has deployed automated classifiers, extensive red-team testing and bug bounty programs to identify and block dangerous requests. Anthropic said testers spent more than 1,000 hours attempting to bypass the model's protections and were unable to find a universal jailbreak.
Security experts told BankInfoSecurity that while Anthropic is temporarily limiting access to Mythos-class cyber capabilities, organizations should use the current generation of AI tools to strengthen defenses now. Yossi Torati, CEO of A Security, said businesses should view today's models as an opportunity to assess vulnerabilities and improve security before more advanced AI-powered cyber capabilities become widely available.
Originally reported by CU Today.