Anthropic has released Claude Fable 5.1, its latest flagship AI model, alongside a specialized version called Mythos 5.1. While Fable 5.1 is broadly available through the company's API, cloud platforms, and desktop application, Mythos 5.1 is restricted to participants in Anthropic's trusted-access program. The latter version features fewer safety restrictions, specifically designed to facilitate research in cybersecurity and the life sciences.
The company claims significant performance improvements for Fable 5.1, particularly in agentic workflows that require multistep problem-solving. Internal benchmarks show the model scoring 55.8% on Terminal-Bench 4.0, an increase from the 42% achieved by its predecessor, Fable 5. In scientific research tasks, the model more than doubled the performance of the previous generation, according to Anthropic's internal testing data.
To support long-running tasks where agents frequently reference the same information, Anthropic has reduced the cost of cache reads by 75%. The price for reading cached input is now $0.25 per million tokens, down from $1.00. While standard API rates for input and output tokens remain unchanged at $10 and $50 per million respectively, the company estimates these pricing adjustments will lower the total cost of typical agentic workloads by approximately 25% to 45%.
The release also marks the implementation of invisible watermarking for text and file outputs, aligning with the EU AI Act. These watermarks, which influence word selection to remain detectable after editing, do not alter the content or quality of responses. Anthropic has also updated its classifiers to reduce false positives, aiming to minimize disruptions for legitimate research work that previously triggered safety blocks.
Looking toward future enterprise needs, Anthropic announced plans to launch Enterprise Frontier Safeguards this fall. This service will allow organizations to maintain monitoring data on their own infrastructure while continuing to utilize Anthropic's misuse-detection systems. This represents a shift from the company's previous requirement that data retention occur on its own servers to manage cybersecurity risks.