Claude Fable 5: Anthropic releases a ‘safe’ version of the Claude Mythos

Anthropic has released Claude Fable 5, a publicly available version of its powerful but previously banned Mythos model – complete with a new set of protections designed to keep its most dangerous power out of the wrong hands. Along with this “safe for general use” model, Anthropic has also released Claude Mythos 5, a version of Fable without security guards, to trusted test partners.

Earlier this year, Anthropic announced the limited launch of the Claude Mythos, a new model with advanced cybersecurity capabilities that Anthropic deemed too risky to release.

The company says the Fable 5 is the most capable model it’s ever made available, leading nearly every benchmark tested across software engineering, information technology, vision, and scientific research. When the task is more demanding, says Anthropic, the Fable 5 has a wide edge over its predecessors and competitors.

Fable 5 shares the same basic architecture as Claude Mythos 5 – the limited version shared with our cybersecurity partners through Project Glasswing – but is shipped in stages that catch sensitive questions and deliver them to Claude Opus 4.8 instead. Restricted categories include cybersecurity, biology, and chemistry, as well as efforts to break down model capabilities for use in competing systems.

Anthropic says less than 5 percent of sessions result in a backlog, though it admits the system is carefully tuned and will sometimes flag wrong requests.

How to try Claude Fable 5

Fable 5 is available today for all Claude programs and through the API using the claude-fable-5 thread model. It is sold at $10 per billion input tokens and $50 per million output tokens – less than half the cost of the Claude Mythos preview. Subscription plan users get access at no additional cost until June 22, after which usage credits will be required.

Measurements

In agent code testing, Fable 5 outperformed GPT-5.5 and Claude Opus 4.8 by significant margins, according to Anthropic. The company’s data shows that it surpasses even Claude Mythos in some important benchmarks.

In a blog post, Anthropic wrote that fintech company Stripe, which had immediate access to Fable 5, reported that the model completed a full migration of the 50-million-line Ruby codebase in one day. Anthropic estimated that the project would have taken a full engineering team more than two months.

Legend 5, Legends 5, and safety

The security issue here is really complicated. Anthropic spent months warning that the Mythos-class models were too dangerous to release. As recently as May, the company publicly acknowledged that adequate safeguards are not yet in place, according to a previous report by Mashable.

The Legend 5 is its answer to that problem, but the company’s disclosures suggest that a solution is still in the works. An external bug bounty has run more than 1,000 hours of testing without producing a global jailbreak – but the UK AI Safety Institute has quickly moved in towards one in a small window to start. Anthropic structures as an acceptable risk. Others may disagree.

The Fable 5 system card states that the model has similar performance to Claude Opus 4.8 and other recent models in non-specific behaviors such as hallucinations, dishonesty, and sycophancy.