Anthropic has launched Claude Fable 5.1 and Mythos 5.1, reducing cache read costs by 75%.

      Anthropic has introduced Claude Fable 5.1 and Mythos 5.1, which are identical models featuring varying safeguards. The company has reduced cache read costs by 75% and reported a score of 52.6% on a scientific research benchmark, compared to 24.7% for the previous version. Outputs from the model include a watermark as mandated by the EU AI Act, and a detection API is currently in private preview for qualifying organizations.

      Two new Claude models have been launched by Anthropic. The company noted that Fable 5.1 and Mythos 5.1 share the same underlying model but differ in their safeguards. Only Mythos is available to the general public, while access to it is granted to selected users through two programs focused on cybersecurity and life sciences, developed in collaboration with the US government.

      A key highlight for developers is the pricing. Cache read costs have dropped 75% to $0.25 per million tokens, reducing typical workload expenses by about 25% and agentic costs by as much as 45%.

      The advancements in benchmarks are significant, according to Anthropic's own data. Fable 5.1 achieves a score of 52.6% on Terminal-Bench-Science 0.1 versus 24.7% for Fable 5, and it scores 55.8% on Terminal-Bench 4.0.

      At the end of the announcement, there’s a mention of the European aspect. In July, Anthropic signed the EU’s Code of Practice regarding the transparency of AI-generated content, along with 190 other signatories. This is the reason Claude's outputs include a watermark, which must be visible only with the detection tool and does not reveal any user information.

      Previously, TNW reported that Anthropic intended to offer a free detection API for verifying text. However, what has been released is more limited; it's currently in private preview for qualified entities such as regulators, law enforcement, media, fact-checkers, independent researchers, educational institutions, and EU civil society groups.

      The requirements set forth are broad. Article 50 mandates that providers must label synthetic outputs in a machine-readable format and ensure their detectability, effective since 2 August 2026. Enterprises with their own compliance responsibilities can also apply, and Anthropic plans to widen access over time, although no specific timeline has been provided.

      Another gap concerning European access pertains to the Mythos model. The access programs have been established with the US government and are exclusive to certain US organizations, two years following the launch of Claude in the EU. Anthropic has stated that it is working to include international partners, although European cybersecurity professionals and life scientists are currently not part of the program.

      This outlines the nature of the launch: the watermark exists due to European legislation, the verification tool is in preview, and TNW has reported that tools for removal are already available.