What Fable 5 and Mythos 5 Are—and Why They Matter
Anthropic’s Claude Fable 5 release and the restricted Anthropic Mythos model are two configurations of the same advanced AI system, designed to test how powerful models can be made widely useful while limiting cybersecurity and biosecurity risks that come with unrestricted AI capabilities. Fable 5 is the public, Mythos-class model: it runs on the same underlying system as Mythos 5 but adds safety gates to limit dangerous outputs. Anthropic describes it as its highest-performing generally available model, with strong results in software engineering, knowledge work, vision and scientific research, especially for long, complex tasks. Mythos 5, by contrast, is distributed only to selected cyberdefenders, infrastructure providers, and life sciences researchers under Project Glasswing, and has many safety restrictions removed so experts can see its full cybersecurity and scientific power. This split is Anthropic’s experiment in balancing AI safety restrictions with demand for frontier performance.
Why Mythos Stays Restricted: Cybersecurity and Dual-Use Power
Anthropic Mythos 5 emerged from Project Glasswing testing that exposed both its promise and its risk. Since an earlier Mythos preview was released in April to about 150 organizations, users reported more than 10,000 critical security flaws in their own systems, showing that the model can find weaknesses faster than many current tools. The same skills could help attackers. Anthropic says Mythos 5 has stronger cybersecurity capabilities than any existing model, so it is limiting access to a small group of cyberdefenders, infrastructure operators and selected biology researchers, in coordination with government agencies. Access is granted on a need-to-know basis, with plans for a broader Trusted Access Program later. This controlled rollout reflects concern that unrestricted, offense-capable AI could undermine digital infrastructure long before defenders adjust, turning a research breakthrough into a security liability.
Fable 5: Mythos-Class Power with Safety Gates and Redirects
Claude Fable 5 is Anthropic’s attempt to put Mythos-level capability in public hands while enforcing AI safety restrictions. Technically, it runs the same underlying model as Mythos 5 but adds a stack of safety classifiers that watch for dangerous topics and jailbreak attempts. When prompts touch sensitive cybersecurity, biology, chemistry or distillation-related queries, Fable 5 does not answer directly. Instead, Anthropic silently routes those requests to Claude Opus 4.8, which returns a safer response while hiding Mythos-class detail. According to Anthropic, fewer than 5% of sessions on average are redirected, though some harmless queries will be caught until the filters improve. The system also monitors for attempts to extract large volumes of answers to train copycat models, sending those to Opus 4.8 as well. For everything outside these restricted categories, Fable 5 is meant to deliver the same performance as Mythos.
Long-Horizon Tasks and the Push into Life Sciences
Both Fable 5 and Mythos 5 are designed for long-horizon tasks, where models plan, code or analyze over many steps instead of responding in short bursts. Anthropic says their advantages become clearer as tasks grow longer and more complex, including autonomous coding workflows and multi-stage research. In life sciences, Mythos 5 already shows how frontier capability can be both helpful and risky. Internal benchmarks cited by Anthropic say Mythos outperformed existing protein language models on adeno-associated virus (AAV) design using biological reasoning alone, without specialized training, and sped up parts of drug design by about tenfold. These same skills could, in principle, help design harmful biological agents. Anthropic’s answer is another restricted track: a biology and chemistry version of Project Glasswing, with Mythos access for selected institutions and, later, verified researchers through a Trusted Access Program instead of open release.
What the Split Strategy Reveals About AI’s Next Phase
The split between Fable 5 and Mythos 5 highlights a turning point in AI development: the most capable systems may no longer be the ones the public uses. Anthropic is experimenting with one frontier model, two exposure levels. Mythos 5, with fewer protections, stays in the hands of trusted partners who are supposed to strengthen security and advance research. Fable 5 offers similar performance for everyday work, but with restricted AI capabilities where abuse risk is highest and with an older model standing in when prompts cross the line. Diane Penn, Anthropic’s head of product management, says this approach “emerged as the most viable and the best one” after months of testing and feedback. For now, Anthropic is choosing to over-block rather than under-protect, signaling an industry shift: future AI releases may be defined less by what models can do than by how much of that power companies are willing to expose.
