Anthropic Launches Fable and Mythos 5.1 Models with Lower Cost and Lighter Restrictions

Anthropic has continued its race in the field of artificial intelligence by announcing the launch of two twin models, Fable and Mythos 5.1, its most advanced releases yet. The new update is not limited to improving performance, but also includes fundamental modifications aimed at reducing the cost of token consumption and minimizing false restriction cases resulting from the model's protection systems.

Two Models with Two Different Paths

Anthropic followed its previous approach in distinguishing between the two versions in terms of access availability. The Mythos 5.1 model will remain exclusive to partners registered with the company, specifically those working in the fields of cybersecurity or life sciences research. As for the unrestricted version, Fable 5.1, it became available from launch day through cloud platforms or through the company's Application Programming Interface (API).

Data Privacy at the Forefront of Concerns

Among the most prominent changes announced by Anthropic is its adoption of the principle of "Zero Data Retention," which allows customers to run the company's models on their own infrastructure without any data leaking to the outside. This feature was not previously available in the Fable model due to security concerns.

During the fall, the company will roll out a high-privacy service called "Enterprise Frontier Safeguards." While the system will continue to monitor any misuse, whether from automated agents or human users, customers will be able to control how this monitoring is carried out.

Anthropic reassured its customers that their data was not used inappropriately, affirming that it "has not trained its models on enterprise data without explicit permission, and will never do so."

Record Performance and Scientific Discoveries

As is customary with Anthropic releases, the two new models set record numbers in a range of benchmark performance tests, most notably:

• Terminal-Bench 4.0: Specialized in programming through the command-line interface.
• Humanity's Last Exam: Dedicated to measuring general reasoning capabilities.

In addition, the company revealed three new scientific discoveries that the models were able to produce before their launch, including a dedicated improvement to GPU (graphics processing unit) performance, as well as the creation of a high-resolution map of Venus compiled from existing images.

System Card and Risk Assessment

Anthropic included a detailed "system card" with the new release, explaining the models' capabilities in a straightforward manner. The card classified the Mythos model as "low risk" regarding concerns about AI self-development, i.e., the model's ability to improve itself, which some experts consider a turning point that could lead to a loss of human control. The card noted that "its ability to accelerate the progress of internal AI research and development is in line with current trends."

Notes on Behavior

Regarding general behavior, the company clarified that Mythos is more prone to behavioral deviation compared to the Opus model, perhaps as a result of its advanced capabilities. The system card stated that Mythos 5.1 represents a slight regression in misaligned behavior compared to Opus 5, but it outperforms both Mythos 5 and Claude Sonnet 5.

The card added that the model "cooperates with human misuse attempts and accepts unverifiable authorization claims more readily than Opus 5," but in return it is "less prone to ignoring explicit restrictions, fabricating fake inputs, or falsely claiming to have completed tasks compared to previous models."

These releases reflect Anthropic's ongoing attempt to balance enhancing capabilities and reducing costs on one hand, while maintaining safety and privacy standards on the other—a challenge facing all leading companies in this rapidly evolving sector.

✦ بقلم فريق دروب أيديا