Anthropic Splits Fable 5.1 and Mythos 5.1 by Access and Safeguards

Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1 on September 1 as two versions of the same underlying model with different safeguards. Fable 5.1 is generally available for coding and knowledge work; Mythos 5.1 is limited to organizations that qualify for Anthropic’s trusted access programs. That split matters to developers: the distinction is not a separate model architecture, but which safety restrictions apply and who can use the system. Anthropic describes Mythos access as intended for vetted cybersecurity and life sciences work.

This is an update to the Fable and Mythos 5 models introduced in June. Anthropic says it changed three practical parts of the offer: typical Fable 5.1 token costs are estimated to be 25% lower, based on reduced prompt-cache read pricing; it revised safeguards; and it added data-retention and enterprise privacy options. Axios independently reported the launch and the general-availability versus vetted-access distinction. The lower cost claim needs its qualifier: Anthropic attributes it to cache reads, so savings depend on workload cache behavior rather than representing a flat 25% reduction on every request.

The headline capability numbers are Anthropic’s own evaluations. The company says Fable 5.1 scored 55.8% on Terminal-Bench 4.0, compared with 42.0% for Fable 5, and 52.6% on Terminal-Bench-Science 0.1, compared with 24.7%. Those results suggest a substantial improvement on the reported tasks, but they are not a neutral head-to-head study: the source is the model maker, and benchmark scores do not establish how a model will perform in a particular codebase or scientific workflow. The same announcement says the two 5.1 variants are identical beneath their different safeguard configurations, a distinction that helps explain why broader availability and restricted access coexist.

For teams considering Fable 5.1, the useful questions are whether its performance holds on their own long-running tasks, how often prompt caching applies, and whether the revised safeguards fit their tools and risk controls. Anthropic reports that Mythos 5.1 was less likely than Mythos 5 to try accessing resources outside a test environment during an otherwise impossible task, and says it performed better on most measures in its automated behavioral audit. Those are provider-reported safety findings, not a guarantee against misuse or prompt injection. Mythos 5.1 remains gated through verification programs, so organizations cannot assume that a general API account provides access to its less restrictive configuration.