What Anthropic’s Two-Tier Claude Model Actually Is
Anthropic’s two-tier Claude approach is a deployment strategy where a single powerful AI model is released in two versions, one broadly available and heavily guarded, and another restricted to vetted defenders with fewer safety limits on sensitive tasks such as cybersecurity. On June 9, Anthropic launched Claude Fable 5 as its most capable model available to the public, while keeping its twin, Claude Mythos 5, behind restricted access for select cyber defenders and critical infrastructure operators. Technically, Fable 5 and Mythos 5 share the same underlying Mythos-class model, but Fable 5 is wrapped in strict AI security guardrails that filter or downgrade high-risk requests. This design lets Anthropic offer near–state-of-the-art performance to most users while reserving its most dangerous capabilities for organizations that pass additional vetting, formalizing a split between general-purpose use and high-stakes security operations.
Inside the Claude Fable 5 Release and Mythos AI Restricted Access
The Claude Fable 5 release marks the first time a Mythos-class model has been opened to the public, offering benchmark performance that Anthropic says matches Mythos on tasks such as SWE-Bench Pro. In practice, Anthropic has shipped one model as two products: Fable 5 for broad use and Mythos 5 with some safety limits lifted for vetted cyber defenders. The company describes Mythos 5 as its strongest cybersecurity model, capable of finding and exploiting software vulnerabilities at a level that could give attackers “serious uplift” if left unchecked. Both Fable 5 and Mythos 5 are priced at USD 10 (approx. RM46) per million input tokens and USD 50 (approx. RM230) per million output tokens, less than half the cost of the earlier Mythos Preview, signaling Anthropic’s intent to make near-frontier capability more accessible while still keeping Mythos AI restricted access for high-trust environments.
How Fable 5’s AI Security Guardrails Work
Anthropic’s tiered model strategy hinges on AI security guardrails built around Fable 5. A set of classifiers continuously monitors prompts for sensitive cyber, biology, chemistry, and model-distillation content. When a request is flagged, Fable 5 does not block outright; it silently routes the response to the weaker Claude Opus 4.8 and informs the user that a fallback occurred. According to Anthropic, this safeguard layer triggers in under 5% of sessions, which means more than 95% of the time Fable 5 behaves like the unrestricted Mythos 5 for ordinary use. In internal tests with blocking enabled, the cybersecurity classifier stopped the model from making progress on offensive tasks such as reconnaissance, lateral movement, and defense evasion. External evaluations also reported zero harmful single-turn responses across 30 public jailbreak techniques, though Anthropic concedes the conservative tuning can cause some false positives that it plans to narrow over time.
Mythos 5: Power for Defenders, Not Attackers
Mythos 5 is the same underlying model as Fable 5, but with key safety switches loosened for approved cyber defenders and security teams. It grew out of the Mythos Preview, which Anthropic initially described as too dangerous to release broadly and limited to roughly 50, then 150, Project Glasswing partners. With Mythos 5, those partners gain direct access to the offensive and defensive cyber capabilities that Fable 5 downgrades, including exploit discovery and complex, long-horizon security tasks. Forrester notes that this aligns with its AEGIS framework principle of least-agency, because Anthropic, not the enterprise, controls the provider-layer guardrails. Enterprises using Mythos still need their own runtime security, but the baseline risk posture is set upstream by Anthropic. The result is a widening gap: organizations that can secure approved access to Mythos gain cutting-edge defensive tools, while others must rely on Fable 5’s safer but more restricted behavior.
Data Retention, Vendor Risk, and the Future of Tiered AI Access
Alongside the new models, Anthropic has shifted its data retention policy in ways that materially affect enterprises. All traffic to Fable 5 and Mythos 5—prompts and completions—now has a mandatory 30-day retention window, even on third-party platforms, overriding any zero-retention agreements tied to earlier contracts. Anthropic says retained data is not used to train new Claude models and is limited to safety uses such as detecting novel attacks, multi-request abuse, and new jailbreak patterns, with human access logged and most data deleted after 30 days. This policy change increases vendor risk considerations, since “safety monitoring” sits next to potential government visibility within the broader regulatory landscape. Strategically, the two-tier design signals a new access model: capability is decoupled from availability, and advanced features depend not just on API keys but on organizational vetting, governance maturity, and an appetite for stricter data-handling terms.
