Anthropic’s Latest Fable Release: More Affordable and Less Restrictive

On Tuesday, Anthropic unveiled Fable and Mythos 5.1, the latest iterations of the company’s most advanced AI model. Along with performance enhancements, the new Fable release brings modifications intended to lower token costs and alleviate false-positive limitations imposed by the model’s protective measures.

Similar to the earlier Mythos model, Mythos 5.1 will be accessible only to registered Anthropic partners involved in cybersecurity or life sciences research. The unrestricted Fable 5.1 is available starting today on cloud platforms or via the Anthropic API.

A key update is Anthropic’s previously announced commitment to zero data retention, enabling clients to utilize Anthropic models on their own infrastructure without any data outflows. Previously unavailable for Fable due to security issues, a high-privacy offering (dubbed Enterprise Frontier Safeguards) will be introduced to users this fall. Significantly, the system will continue to track any misuse by agents or human operators, but clients will have control over how this monitoring is conducted.

In the announcement, Anthropic reassured customers that their data had not been improperly accessed. “Anthropic has never trained on enterprise data without explicit permission, and will never do so,” the statement declares.

As is typical with an Anthropic launch, the new models achieve records across various benchmarks, including Terminal-Bench 4.0 (for CLI-based coding) and Humanity’s Last Exam (for general reasoning). Anthropic also shared three unique scientific discoveries generated by the models prior to their release, including a tailored GPU optimization and a high-resolution map of Venus constructed from existing images.

In line with past releases, the models come with a comprehensive system card, which lays out their capabilities in clear terms. The system card evaluates Mythos as “low-risk” concerning automated AI development—where AI enhances itself—something that some regard as a potential loss of human oversight. It indicates that “its capacity to accelerate internal AI R&D progress aligns with ongoing trends.”

Regarding general misbehavior, Mythos is slightly more susceptible than Opus, likely due to its advanced capabilities.

“Mythos 5.1 represents a minor regression in overall misaligned behavior compared to Opus 5, yet shows improvement over Mythos 5 and Claude Sonnet 5,” the system card states. “It collaborates with human misuse and accepts unverifiable assertions of authorization a little more easily than Opus 5, but it is less prone to ignore clear limitations, hallucinate inputs, or falsely claim task completion than earlier models.”