Anthropic has released two updated versions of its flagship AI systems — Fable 5.1 and Mythos 5.1. According to the developers, both models have become more performant, cheaper to operate, and exhibit fewer false positives from their built-in safety systems. The release is accompanied by a revised data-handling policy and an updated system card in which the company openly assesses the behavioral risks of the new versions.
Two Models, One Release
The update affected both product lines: Fable 5.1 and Mythos 5.1 gained a performance boost and lower per-call costs, making them more accessible to mainstream developers. At the same time, Anthropic has relaxed the behavior of its safety filters, which in previous versions frequently blocked legitimate requests. This is directly reflected in independent reviews, which note the more "lenient" behavior of the new models while basic protective mechanisms remain in place.
Zero Data Retention and Enterprise Control
The main infrastructure innovation of this release is Anthropic's shift to a zero data retention policy. Clients can now run the models on their own infrastructure without the risk of data leaking to the company's external servers. According to industry publications, this decision was made following active pressure from enterprise customers, for whom data confidentiality was a key barrier to AI adoption. In the fall, users will additionally receive the Enterprise Frontier Safeguards service — high-frequency monitoring of AI actions under full client control. The developers also assured that companies' private data has never been used and will not be used to train models without explicit permission.
Benchmarks and Independent Discoveries
Both systems posted record results on Terminal-Bench 4.0, which evaluates command-line programming, and on the complex, comprehensive Humanity's Last Exam task suite. Notably, even before the official release, the models completed a number of independent tasks: they optimized GPU architecture and built a highly detailed map of Venus based on available photographs. These examples underscore the growth of agentic capabilities — the ability of AI to independently break down and execute multi-step tasks.
Separate Distribution Policy
Access to the models remains segmented. Fable 5.1 is available to all developers, whereas the Mythos 5.1 version will be provided by Anthropic exclusively to registered partners engaged in cybersecurity or life-sciences research. This approach allows the company to control the deployment of its most powerful model in sensitive domains.
System Card and Risk Assessment
Alongside the release, the company published a system card with an analysis of the AI's behavior. Experts assessed Mythos as a "low-risk" model with respect to scenarios of loss of human control due to autonomous self-improvement. At the same time, Mythos 5.1 shows a somewhat greater tendency toward incorrect behavior than the Opus 5 version, but outperforms Mythos 5 and Claude Sonnet 5 in reliability. Improved instruction-following is noted separately: the model less often ignores direct instructions, less often fabricates input data, and does not present fabricated task completion as real.
