What Anthropic’s Split Claude Models Are and Why They Matter
Anthropic’s split Claude model strategy is a two-tier design where one highly capable AI system is offered in a safety-gated public version while the same core model, with fewer restrictions, is reserved for vetted security professionals to limit misuse while still providing strong defensive tools. With the Claude Fable 5 release, Anthropic made its most capable model generally available, but paired it with Claude Mythos 5, a twin version kept for selected cyber defenders and critical infrastructure operators. Both share the same underlying model, yet differ through safety classifiers that detect and reshape high-risk outputs. This is not a rollback but a deliberate structure: Mythos 5 keeps advanced cybersecurity abilities exposed for trusted users, while Fable 5 routes some sensitive prompts to an older, weaker model, Claude Opus 4.8, to reduce the risk that attackers gain significant uplift.
How Anthropic Safety Gating Works in Claude Fable 5
Anthropic safety gating in Claude Fable 5 relies on separate AI classifiers that watch incoming queries for signs of misuse or jailbreak attempts. When a request touches sensitive areas such as cyber operations, biology, chemistry, or model distillation, Fable 5 does not refuse outright. Instead, it hands the response to Claude Opus 4.8, and the interface tells the user that a fallback occurred. According to Anthropic, fewer than 5% of all sessions trigger this redirection, which means that in more than 95% of use, Fable 5 behaves like the unrestricted Mythos 5. Early tests show the safeguards can withstand a range of jailbreak methods, including 30 public techniques in one external evaluation. Anthropic admits false positives are an issue today but says it plans to narrow the classifiers so that harmless queries are less likely to be interrupted over time.
Mythos 5: Stronger Cyber Capabilities for Vetted Defenders
Claude Mythos 5 is Anthropic’s higher-capability tier, designed as an upgraded version of the earlier Mythos preview and distributed only through a trusted access program. Anthropic calls Mythos 5 “the strongest cybersecurity model in the world,” built on the same base as Fable 5 but without the public safety gates that redirect offensive cyber, bio, or chemistry tasks. Earlier testing of the Mythos preview showed the risks: the model identified and exploited zero-day vulnerabilities across major operating systems and browsers and even wrote a remote code execution exploit against a long-standing FreeBSD NFS bug. Those demonstrations are why Mythos 5 is now limited to selected security experts and critical infrastructure providers. Anthropic says it will gradually widen access, but only for organisations that can use these offensive capabilities for defense, not opportunistic exploitation.
AI Capability Tiers as a New Safety Tradeoff
The Claude Fable 5 release signals a new way of handling the tension between AI progress and safety. Instead of weakening a model before public launch, Anthropic maintains full capability in Mythos 5 and then builds Anthropic safety gating around Fable 5 to create AI capability tiers. For everyday users, Fable 5 delivers Mythos-level performance on software engineering, knowledge work, vision and scientific tasks, while quietly defusing prompts that might expose frontier cyber or life sciences know-how. For vetted defenders, Mythos 5 keeps advanced exploit discovery and attack-chain reasoning available. This tiered strategy acknowledges that the same tool can patch systems or break them, depending on who holds it. If it works, it could become a template: strong models shipped once, then segmented by access controls and classifiers instead of by permanently cutting their abilities.
Economic and Scientific Stakes of the Two-Tier Claude Design
Anthropic has also tied this safety-first structure to its business and research roadmap. Claude Fable 5 and Mythos 5 share the same pricing, at USD 10 (approx. RM46) per 1 million input tokens and USD 50 (approx. RM230) per 1 million output tokens, and Fable 5 is temporarily included on paid plans at no extra cost before moving to usage credits. That equal pricing suggests the split is about risk, not upselling capability. On the science side, Anthropic is planning a Project Glasswing track focused on life sciences and biochemistry, where the same patterns of dual use apply: the ability to design therapies sits close to the ability to design threats. Mythos-style models for biology will likely follow the same path, with Mythos model safety limits for the public and powerful versions reserved for carefully screened research institutions.
