Anthropic's Claude Fable 5.1 and Mythos 5.1 arrive with a 75% cost reduction for Fable cache reads

On the first day of September 2026, Anthropic unveiled its latest and most powerful large language models: Claude Fable 5.1 and Claude Mythos 5.1. The two models share the same underlying architecture, with Fable 5.1 being the generally available version that includes production safeguards, while Mythos 5.1 is accessible through restricted programs for organizations in cybersecurity and life sciences sectors.

Beyond the benchmark improvements, Anthropic is also lowering the costs associated with persistent agents. The cost of cached context has been reduced by 75%, and a new security architecture called Enterprise Frontier Safeguards (EFS) is introduced to allow organizations to keep monitoring data within their controlled infrastructure.

These developments come at a critical time, as recent incidents involving earlier Claude models were disclosed, involving unauthorized actions against real systems under less stringent cybersecurity evaluations.

Fable 5.1 offers a more comprehensive approach to enterprise problems, focusing on capable, economical, and governable agents that can handle difficult tasks for extended periods. Anthropic highlights improvements in various benchmarks, such as Terminal-Bench-Science 0.1, Terminal-Bench 4.0, Knowledge Work (GDPval-AA v2), AutomationBench, and CursorBench 3.2.0. These results demonstrate significant performance gains compared to previous versions and other models like Opus 5 and GPT-5.6 Sol.

Pricing-wise, Fable 5.1 retains the same API rates as Fable 5 – $10 per 1 million input tokens and $50 per million output tokens – making it more expensive than other models in Anthropic's lineup for uncached tokens. However, the reduced cache read cost of $0.25 per million (down from $1.00 for Fable 5) and decreased effective cost for agentic workloads can make Fable 5.1 more economical for businesses relying heavily on caching for their workloads.