Anthropic launches Claude Fable 5.1 with up to 45% lower agentic task costs

Anthropic announced yesterday the release of Claude Fable 5.1 and Mythos 5.1, two updates that directly address customer feedback on pricing, data retention and over-protective safeguards. Claude Fable 5.1 delivers performance comparable to its predecessor, while the average cost has dropped by about 25 percent, and for complex agentic tasks the discount can reach up to 45 percent thanks to reduced pricing for cached data after an initial processing pass.

The cost reduction stems from a change in how the company prices reuse of already-processed context. In scenarios where an agent repeatedly accesses the same material—such as large code-base analysis or long-document processing—the cumulative expense is now significantly lower than before. Mythos 5.1 is being launched concurrently as part of Project Glasswing, a closed early-access program that is not yet available to the broader user base.

Early impressions from beta users highlight improvements in coding and natural-language capabilities. Dan Shipper, CEO of Every, who received early access, described Claude Fable 5.1 as “the strongest coding model our team has used,” adding that it is fast, token-efficient, and “talks like a normal human.” Aaron Levie, CEO of Box, reported that a Claude Fable-based agent detected nuances and ambiguities in data that the previous version missed in the same test. Lisan al Gaib noted that, with a low-reasoning setting, Mythos 5.1 achieves the same result as its predecessor on maximum-reasoning configurations.

Claude Fable 5.1 also includes “more precise safeguards,” according to Anthropic, reducing the likelihood that the model will block basic biology questions. Mythos 5.1 retains the same biological limitations as the prior generation. On the privacy front, the company announced that “Enterprise Frontier Safeguards” will provide “full privacy” by storing data on the customer’s own cloud servers rather than Anthropic’s infrastructure, with a phased rollout planned for the coming fall.

Anthropic now permits Claude Fable 5.1 to be used for software vulnerability identification, but more aggressive activities—penetration testing, exploit generation and binary-based vulnerability scanning—remain directed to Opus models. Claude Fable 5.1 is available now across all platforms; Mythos 5.1 remains limited to Project Glasswing participants.