Anthropic has released the new generation of models Fable 5.1 and Mythos 5.1. Both models share the same generation of capability upgrades, but the scope of availability is different. Fable 5.1 has been launched through the cloud platform and on Anthropic API, while Mythos 5.1 continues to be available only to specific partners.
Fable is now open to a wider range of customers.
One of the key focuses of this update is to lower the entry barrier. Anthropic indicates that in Fable 5.1, the token cost has been reduced, and at the same time, there has been a decrease in false positives in security protections, resulting in fewer cases where legitimate requests are incorrectly denied.
In contrast, Mythos 5.1 is still only available to registered partners, with its use limited to network security and life science research. Anthropic continues to maintain a "dual version" strategy, placing more sensitive capabilities in controlled scenarios.
Enterprise-level high-privacy services will be launched in June.
Anthropic also advanced the previously mentioned Zero Data Retention initiative. According to the company, customers can run models on their own infrastructure, avoiding data leakage. This capability was not previously available to Fable due to security concerns.
Now, Anthropic plans to launch a high-privacy service named Enterprise Frontier Safeguards for Fable, which will be available to users in June. The system will still monitor for any abuse by proxies or human users, but the method of monitoring will be under the control of the customers themselves.
The company also stated that no corporate data was used for training without explicit permission, and there were no instances of unauthorized access to corporate data.
Benchmark scores refreshed, with varying levels of security performance
According to Anthropic, the new model has refreshed its performance in multiple benchmark tests, including Terminal-Bench 4.0, which assesses command-line programming capabilities, as well as Last Exam of Humanity, which tests general reasoning abilities.
The company also stated that before its release, the model generated three new scientific research achievements, including a customized GPU optimization scheme and a high-resolution map of Venus created by splicing existing photographs.
In terms of security assessment, the system classifies Mythos 5.1 as having low risk regarding automated AI development-related risks. However, Anthropic also acknowledges that the model has seen a slight regression in some indicators of improper behavior compared to Opus 5, but there have been improvements in areas such as ignoring clear restrictions, fabricating input content, and falsely claiming task completion compared to previous versions.
