Anthropic makes a new “Mythos-class” model available to the general public and says it includes safety guardrails for higher-risk requests. According to reporting, when users ask questions involving potentially dangerous areas—such as cybersecurity or bioweapons—the system does not answer directly with the new model. Instead, the requests are steered to an older Opus model. The approach is intended to manage behavior around topics that can enable harm, while still allowing users to interact with the assistant for legitimate questions. The published descriptions emphasize that the new release is about making the model broadly accessible and adding a routing mechanism for certain categories of risky prompts. Across outlets, the central details are the availability of the Mythos-class model and the use of model switching—sending specific dangerous-topic queries to Opus rather than the new system. Other operational details about the guardrails’ exact triggers, scope, or performance trade-offs are not fully specified in the provided excerpts.