Claude Fable 5.1 and Claude Mythos 5.1

Anthropic has unveiled Claude Fable 5.1 and Claude Mythos 5.1, two of the most advanced models designed for coding and knowledge work. The new models boast remarkable performance improvements and specialized research capabilities, opening a window into how AI could support scientific progress.

Fable 5.1 is now generally available, while Mythos 5.1 remains accessible through Anthropic's trusted access programs. The latter's safeguards are tailored for cybersecurity and life sciences applications. Despite the enhanced features, Fable 5.1 offers notable cost savings, with token-based pricing slashed by 25% compared to Fable 5, and even larger reductions of up to approximately 45% for highly agentic tasks. This price reduction stems from Anthropic's efforts to cut cache read costs and storage expenses.

In response to customer feedback, Anthropic has implemented new safeguards to minimize false positives and bolster data privacy. The Enterprise Frontier Safeguards (EFS) enable customers to maintain complete privacy akin to a zero data retention policy by storing data in customer-controlled cloud infrastructure. This feature is set to roll out to enterprise customers later in the fall, with eligible users able to employ Fable 5.1 with zero data retention until then.

Anthropic's enhancements to Fable 5.1 include reduced false positives in cybersecurity applications, blocking 60% fewer false positives than before. In biology, their partnerships with the US government facilitate research using Claude Mythos 5.1's advanced biology capabilities, with early access programs planned for scientists.

Claude Fable 5.1 sets a new benchmark for coding and long-running problem-solving tasks. Benchmarks demonstrate exceptional performance gains compared to its predecessor, Fable 5. In testing by investment firm Millennium, Fable 5.1 identified the cause of a persistent crash in internal systems, a mystery that eluded their engineers for years.

The model outperformed competitors in molecular design tasks, crafting high-affinity binders with binding affinities 10 times greater than the best human designs. In computational biology, it swiftly trained a neural network that optimized seven deep learning models, achieving up to 2.5 times faster performance, resulting in estimated GPU cost reductions of 30–60%.