Anthropic launches Claude Sonnet 5 and restores Fable and Mythos models following safety upgrades
Anthropic has released the Claude Sonnet 5 model and restored access to its Fable and Mythos systems following the implementation of new government-mandated safety protocols.
1. Restoration of Frontier Models
On July 1, 2026, Anthropic resumed access to its Fable and Mythos frontier AI models following an eighteen-day operational suspension. The shutdown was mandated by a U.S. government export control directive issued on June 12, 2026, after Amazon researchers discovered a method to bypass safety controls in Fable 5, allowing the model to generate software exploitation code. To resolve the federal directive, Anthropic engineers developed an automated safety classifier that blocks prompts with a high statistical probability of malicious intent. Internal data indicates this patch prevents the identified exploitation technique in over 99 percent of trials.
2. Launch of Claude Sonnet 5
Alongside the restoration of its frontier models, Anthropic launched Claude Sonnet 5. Designed for enterprise use, the model is being integrated into autonomous agent workflows to handle multi-step planning, terminal operations, and web navigation. Performance benchmarks indicate that Sonnet 5 outperforms its predecessor, Sonnet 4.6, in both coding and terminal-based tasks. Early adopters, including Rakuten, Zapier, and Zed, have utilized the architecture to automate complex software development, debugging, and administrative processes.
3. Governance and Security Frameworks
The regulatory challenges surrounding the Fable 5 suspension have led to a new industry collaboration between Anthropic, Amazon, Microsoft, and Google. These companies are working to establish a standardized framework for assessing AI security breaches, focusing on metrics such as capability gain, weaponization ease, and discoverability. Furthermore, Anthropic has entered into agreements to provide federal researchers with early access to frontier models for security audits prior to public release. While the new safety classifiers have increased the model's robustness, they also introduce a trade-off, as the system may more frequently flag benign requests during routine development tasks.