Anthropic's Fable AI: Cheaper, Freer, and Ready to Disrupt
Anthropic has launched Fable 5.1 and Mythos 5.1, its most advanced AI models, featuring performance upgrades and reduced token costs. Key innovations include a zero data retention policy and a high-privacy Enterprise Frontier Safeguards service, bolstering data security and client control. These models have also set new benchmarks and contributed to novel scientific findings.
Anthropic has officially announced the release of Fable 5.1 and Mythos 5.1, its latest and most advanced artificial intelligence models. These twin versions introduce significant performance enhancements alongside crucial changes aimed at optimizing token cost and reducing false-positive restrictions from the models' integrated safeguards.
Mythos 5.1 is specifically designed for and exclusively available to registered Anthropic partners, focusing on specialized applications within cybersecurity and life sciences research. In contrast, Fable 5.1, the unrestricted counterpart, is now broadly accessible via cloud platforms and through the Anthropic API, allowing a wider range of users to leverage its capabilities.
A notable advancement with this release is Anthropic's adoption of a zero data retention policy. This allows clients to deploy and run Anthropic models on their proprietary infrastructure, eliminating data outflows and enhancing privacy. Furthermore, a new high-privacy service, termed Enterprise Frontier Safeguards, is set to roll out in the fall. Previously unavailable for Fable due to security considerations, this service will enable monitoring for misuse by agents or human users, with clients maintaining control over the monitoring mechanisms. Anthropic has also reiterated its commitment to data privacy, assuring customers that enterprise data has never been, and will never be, used for training without explicit permission.
In line with previous releases, the new models have established new performance benchmarks. They have set records in various evaluations, including Terminal-Bench 4.0 for command-line interface-based coding and Humanity’s Last Exam, which assesses general reasoning abilities. Pre-release, the models also generated three novel scientific findings: a custom GPU optimization and a high-resolution map of Venus, compiled from existing photographic data.
Detailed capabilities of Fable 5.1 and Mythos 5.1 are outlined in a comprehensive system card. This document rates Mythos as "low-risk" concerning automated AI development – a scenario where AI systems improve themselves, which some experts view as a potential trigger for a loss of human control. The system card states that Mythos's capacity to accelerate internal AI research and development progress aligns with current industry trends.
Regarding general misbehavior, the system card notes that Mythos 5.1 exhibits a "slight regression on overall misaligned behavior compared to Opus 5," although it represents an improvement over Mythos 5 and Claude Sonnet 5. It is described as being somewhat more cooperative with human misuse and more accepting of unverifiable claims of authorization than Opus 5. However, it is also highlighted as being less prone to ignoring explicit constraints, hallucinating inputs, or falsely claiming task completion when compared to earlier models.