Anthropic has announced "Claude Fable 5.1," the latest large language model with improved inference and development performance, as well as "Claude Mythos 5.1," which features safety measures tailored for cybersecurity. For Fable 5.1, the price for cache reads—which reuse stored input contexts—has been set at $0.25 per million tokens, a 75% reduction compared to previous pricing. Anthropic stated that this could lead to cost savings of up to approximately 45% depending on the workload.
Updates were also made to safety; improved safeguard precision has reduced the false positive rate where safe prompts are mistakenly rejected. Consequently, the identification of software vulnerabilities is now officially permitted in Fable 5.1, enabling its use for tasks such as code auditing for defensive purposes. Additionally, in compliance with the EU AI Act, invisible watermarks will be embedded in the outputs of models released after August 2026.
Source: 「最大45％コスト減」 Anthropicが「Claude Fable 5.1」発表 脆弱性検知にも対応 - ITmedia (Google News: Anthropic)