Artificial intelligence company Anthropic has rolled out a public version of its advanced Mythos AI model, introducing new safety restrictions that prevent its use in high-risk areas such as cybersecurity.
The company said the updated system, branded Claude Fable 5, is its most powerful model yet for general public use. However, unlike earlier restricted previews, the public release includes strict guardrails designed to block attempts to use the model for identifying or exploiting software vulnerabilities.
Anthropic had previously limited access to around 200 organisations, including select US government entities under a controlled testing programme, after earlier versions of Mythos demonstrated strong capability in discovering software flaws. That capability raised concerns in the cybersecurity community about potential misuse.
With the public rollout, the company said it had implemented additional safeguards to ensure the model cannot be manipulated to bypass restrictions. If a user attempts to request sensitive cybersecurity assistance, the system is designed to refuse the request or redirect it to a less capable fallback model.
Company officials said the goal is to balance advanced performance with responsible deployment. According to Anthropic, Claude Fable 5 offers improved efficiency in software engineering and data analysis tasks while reducing overall computational cost per request.
The firm also noted that the model is designed to use fewer tokens per task, making it more cost-effective for customers despite its higher baseline pricing. The company has set pricing at $10 per million input tokens and $50 per million output tokens.
Users who previously accessed the unguarded preview version will be able to upgrade to the safer public release, while access to more powerful or less restricted versions will be managed through a “trusted access” programme.
Anthropic said it will gradually expand availability as it continues testing and refining safeguards, as major AI developers face increasing pressure to ensure powerful models are deployed responsibly.