Anthropic has unveiled two new versions of its most advanced AI system, Fable 5.1 and Mythos 5.1. In addition to improving performance, the updates aim to reduce token costs and limit false positives related to model protection.
Anthropic has unveiled two new versions of its most advanced AI system, Fable 5.1 and Mythos 5.1. In addition to improving performance, the updates aim to reduce token costs and limit false positives related to model protection.
Regarding the availability of the new models, Fable 5.1 — a version with relaxed restrictions — is now open to all users via cloud platforms and the Anthropic API, while the highly secure Mythos 5.1 is provided exclusively to the company's registered partners engaged in research in the fields of cybersecurity and life sciences, writes TechCrunch.
One of the key changes is the introduction of a zero-data retention concept. This will allow customers to run Anthropic models on their own infrastructure without the risk of data leakage. An enhanced privacy service called Enterprise Frontier Safeguards, which was previously unavailable to Fable due to security concerns, will be available in the fall. The system will continue to monitor for abuse by users or autonomous agents, but customers will have control over the process of this monitoring.
Anthropic assured users that their corporate data was not compromised: “Anthropic has never trained and will never train models on corporate data without explicit permission.”
Both models set new records in a series of popular benchmarks, including Terminal-Bench 4.0 (testing coding via the command line) and Humanity's Last Exam (assessing general reasoning).
Before the model was released, three scientific discoveries were made, including:
special optimization for graphics processors (GPU);
a highly detailed map of Venus, compiled from existing photographs.
According to the published system card, Mythos has a “low risk” for threats related to independent AI development. The model’s ability to accelerate internal AI research and development is in line with current industry trends.
In terms of general behavior, Mythos 5.1 is somewhat more likely to detect anomalies than Opus 5, which is explained by its increased capabilities: “Mythos 5.1 has a slight regression in behavioral anomalies compared to Opus 5, but is an improvement over Mythos 5 and Claude Sonnet 5. It is somewhat more cooperative in case of incorrect user actions or accepts unconfirmed claims of authorization than Opus 5, but is less likely to ignore explicit restrictions, fabricate input, or falsely report task completion.”
