What Claude Mythos is
Claude Mythos is Anthropic's highest-capability model tier, positioned above the Opus/Sonnet/Haiku line and designed for tasks that require frontier-level power in domains — particularly cybersecurity — where that power also creates meaningful risk. It does not replace the existing naming scheme; rather, it sits above it as a distinct product with a distinct access model.
The Mythos line ships as a dual-variant system. Claude Mythos 5 is the full-capability version, available only to vetted partners through Project Glasswing, a program developed in collaboration with the U.S. government for cyberdefense and critical infrastructure work. Claude Fable 5 is the general-availability version of the same underlying model, wrapped in safety classifiers that block or degrade responses on cybersecurity, biology, chemistry, and cutting-edge AI-development topics. Both are priced at $10/$50 per million input/output tokens — roughly half the cost of the prior Claude Mythos Preview.
Capability profile
Both variants set new state-of-the-art results across software engineering, agentic coding, knowledge work, cybersecurity, and scientific reasoning benchmarks at launch. Commentary from Zvi Mowshowitz described Claude Fable 5 as the best publicly available model at release. The predecessor, Claude Mythos Preview, had already attracted attention for solving a novel cybersecurity challenge — a capability that prompted meetings with the Financial Stability Board and, separately, drew unauthorized access attempts via Discord coordination before the formal Mythos 5 launch.
Anthropic previewed the Mythos class above Opus in early June 2026, framing it as pending broader safety clearance. The Andon Labs team, creators of VendingBench, noted they had evaluated Claude models across the full spectrum from Haiku to Mythos in building frontier evals, suggesting the tier had been accessible to select evaluators before general announcement.
The dual-variant architecture and its tensions
The Mythos/Fable split is the most operationally novel aspect of this release. Rather than a single model with a single policy, Anthropic is maintaining two deployment surfaces of the same weights under different constraint regimes — a pattern that introduces structural complexity for users, evaluators, and regulators alike.
Silent degradation. At launch, Fable 5 applied undisclosed capability degradation to AI-development prompts via prompt modification or steering vectors. Users and researchers were not informed that responses in this domain were being silently altered rather than refused or answered normally. After researcher backlash, Anthropic reversed the policy — but the episode raised durable questions about transparency in safety-tiered deployment.
Evaluability breakdown. The safety classifiers created a measurement crisis. Fable 5's classifiers silently rerouted flagged prompts to Claude Opus 4.8 rather than refusing them explicitly. Independent evaluators — Artificial Analysis, Vals AI, and ARC Prize Foundation — each adopted different strategies for handling these reroutes (blended scores, pure scores, or abstaining entirely), producing wildly divergent rankings. On GPQA Diamond alone, Fable 5's score ranged from 93.18% (2nd place) to 55.56% (94th place) depending on how refusals were counted. This is not a minor scoring disagreement; it surfaces a structural incompatibility between safety-oriented deployment constraints and the field's ability to independently measure frontier capability.
Data retention. Anthropic's new data retention policy further complicated independent evaluation, limiting what evaluators could log and analyze about model behavior under classifier conditions.
Regulatory collision
Three days after launch, the U.S. Commerce Department issued an export control directive forcing Anthropic to suspend global access to both Claude Mythos 5 and Claude Fable 5. The suspension — flagged first on Anthropic's status page on June 13, 2026 — represented an unprecedented regulatory intervention: a government-mandated withdrawal of a frontier model almost immediately after release.
Andrew Ng's commentary in The Batch framed the episode as a demonstration of how both private companies and governments can suddenly restrict AI access, and argued it would accelerate global interest in AI sovereignty and open-source alternatives. Over 100 cybersecurity professionals subsequently signed an open letter urging the U.S. government to reverse the export controls, citing the defensive value of Mythos-class models for infrastructure protection.
Competitive context
The Mythos launch prompted a direct competitive response: OpenAI launched GPT-5.4-Cyber, a vetted-access model for defensive cybersecurity tasks, in the same period. The parallel suggests a broader industry shift toward capability-tiered deployment for high-risk domains — where the question is not just what a model can do, but who is allowed to use it and under what conditions.
Structural implications
The Mythos/Fable architecture represents a significant escalation in the operational complexity of frontier model deployment. Managing two access tiers of the same model, maintaining classifier transparency, satisfying independent evaluators, and navigating government export controls simultaneously is a qualitatively different challenge than shipping a single general-purpose model. Whether this dual-variant pattern becomes an industry template — or whether the evaluability and regulatory friction it generates proves unsustainable — is the open question the Mythos launch has placed on the table.
