Anthropic is releasing a public version of its Mythos AI model as Claude Fable 5

Rather than a separate architecture, Fable 5 runs on the same foundation as Mythos 5, with added classifiers that intercept and block outputs touching high-risk domains — cybersecurity, biology and chemistry, and prompts Anthropic identifies as attempts to extract model capabilities for use in rival AI systems. Queries that trip those classifiers are handed off to Claude Opus 4.8, a lower-capability model, rather than answered by Fable 5 directly. According to Anthropic, the handoff to Opus 4.8 is uncommon, with the vast majority of Fable 5 sessions — more than 95%, by early data — handled entirely by Fable 5 itself.

At launch, Anthropic set Fable 5's API rate at $10 per million input tokens and $50 per million output tokens — a significant reduction from the $25 and $125 per million token pricing that had applied to the restricted Mythos Preview since April. Anthropic said it expects demand to be high and will roll out access in stages for subscription plan users. Subscribers on Pro, Max, Team, and seat-based Enterprise plans can use Fable 5 without additional charges through June 22. Beginning June 23, access through those subscription tiers will be suspended and replaced by a credits-based model, with Anthropic intending to bring it back into standard plans once supply catches up with demand.

Separately, Anthropic made Claude Mythos 5 available Tuesday to the set of organizations that had previously cleared its approval process for Mythos access. It is the same model as Fable 5 but with cybersecurity safeguards lifted in some areas. Mythos 5 will initially be deployed through Project Glasswing, in collaboration with the U.S. government, as an upgrade to Claude Mythos Preview.

Before making Fable 5 publicly available, Anthropic subjected its safety classifiers to intensive adversarial testing, including a bug bounty program that logged over 1,000 hours of attempts without any participant finding a universal jailbreak, according to TechCrunch. External red-teaming organizations also failed to find universal jailbreaks on long-form tasks, though Anthropic acknowledged that the U.K. AI Security Institute made progress toward one in an initial testing window.

With the launch, Anthropic is also changing its data retention policy for Mythos-class models. Going forward, all Mythos-class traffic will be subject to a mandatory 30-day retention window — a policy shift that overrides prior zero-retention arrangements some enterprise customers had negotiated — so Anthropic can monitor for emerging attack patterns and cut down on classifier errors. The company specified that training future models is explicitly excluded as a use for the retained data.

Anthropic restricted Mythos to a limited group of cybersecurity partners and critical infrastructure providers when it launched in April, citing concerns about the model's ability to identify and exploit software vulnerabilities. The original Mythos Preview carried a price of $25 per million input tokens and $125 per million output tokens. Partners included JPMorgan $JPM Chase, Apple $AAPL, Amazon $AMZN Web Services, Microsoft $MSFT, Google $GOOGL, and CrowdStrike $CRWD, among others. The model was accessed without authorization by a small group of users on the day of its announcement, a breach Anthropic said it investigated and found contained within a third-party vendor environment.