What Claude Fable 5 Is — And Why Mythos 5 Stays Gated
Claude Fable 5 is Anthropic’s first generally available Mythos-class model, a highly capable AI system that matches its restricted twin, Claude Mythos 5, in core capabilities but adds safety classifiers that redirect sensitive cyber, biological, chemical, and model-distillation requests to a weaker model instead of answering them directly. Anthropic released Fable 5 to the public across its API and major cloud platforms, while keeping Mythos 5 restricted to vetted Project Glasswing partners such as cyber defenders and critical infrastructure operators. Both are built from the same underlying model, but Anthropic describes Mythos 5 as its “strongest cybersecurity model,” with fewer guardrails for approved users. This two-tier setup makes Claude Fable 5 release a turning point: high-end capabilities reach everyday developers, yet the most dangerous uses remain locked behind institutional vetting and oversight rather than open access.
Benchmarks, Guardrails, And The Mythos Model Restrictions
Anthropic’s split is notable because Fable 5 is not a downgraded product. On SWE-Bench Pro, a demanding coding benchmark, Fable 5 scores 80%, while Mythos 5 reaches 80.4%, both ahead of Claude Opus 4.8 at 69.2% and rivals like GPT 5.5 and Gemini 3.1 Pro. In other words, the public model still performs at the frontier despite its AI security safeguards. The difference lies in what happens when queries cross sensitive lines. Fable 5 uses separate classifiers to detect cyberattack planning, exploit development, biological and chemical threat content, and model distillation attempts. When triggered, the system falls back to Opus 4.8 and informs the user about the handoff. Early partner tests reported that, with blocking enabled and no evasion attempts, these classifiers prevented progress on offensive cyber tasks and withstood 30 known jailbreak techniques, trading some false positives for a tighter safety boundary.
How Anthropic Balances Capability, Cyber Risk, And Access
Anthropic’s design choice shows its view of AI governance: powerful models should exist, but their most dangerous capabilities should not be available to everyone by default. Mythos 5 is reserved for vetted organizations that can apply their own stringent controls, while Fable 5 bakes in guardrails for the wider market. According to Forrester, “Fable 5 and Mythos 5 are one model with a safety switch,” with Anthropic alone controlling how sensitive that switch is. This structure widens the gap between well-resourced security teams and everyone else. Project Glasswing members gain access to offensive and defensive cyber features that remain blocked in the public product. At the same time, Fable 5’s state-of-the-art performance and lower price than the earlier Mythos Preview model signal an era where cost optimization and safety tuning become core parts of AI strategy, not fringe concerns for specialists.
User Friction: Usage Limits, Pricing, And Fallback Behavior
For everyday users, the Claude Fable 5 release arrives with both upside and friction. Fable 5 is available through the Claude API and on platforms such as Microsoft Foundry, Amazon Bedrock, and Claude on AWS at USD 10 (approx. RM46) per million input tokens and USD 50 (approx. RM230) per million output tokens, which Anthropic notes is twice the price of its Opus model but less than half the price of the earlier Mythos Preview. Subscribers on Pro, Max, Team, and seat-based Enterprise plans can use Fable 5 at no extra cost until June 22; after that, usage draws from separate credits, introducing budget planning headaches for high-volume teams. Because sensitive prompts can silently fall back to Opus 4.8, enterprises also need to account for changing behavior mid-session, including potential shifts in coding or reasoning style and slightly weaker performance on restricted tasks.
Data Retention Policy And The New Vendor Risk Equation
The most disruptive change for enterprises may be Anthropic’s updated data retention policy. For traffic to Fable 5 and Mythos 5, Anthropic now requires 30-day retention of all prompts and completions across its own products and third-party platforms, overriding any existing zero-retention data processing agreements. There is no opt-out for Mythos-class usage. This shifts AI vendor risk calculations. Security leaders must now assume that sensitive prompts, documents, and generated code will live on Anthropic’s systems for a month, even if prior contracts said otherwise. Combined with automatic fallback to Opus 4.8 for certain topics, this means data may be processed by multiple models and stored under a unified policy. Organizations that want Mythos model restrictions lifted for advanced cyber defense now have to weigh that benefit against stricter internal governance, tighter access controls, and updated risk disclosures to boards and regulators.
