Anthropic released Claude Fable 5, a safety-focused AI model, to the public this week [1].
The launch represents a strategic attempt to balance the deployment of powerful AI capabilities with the prevention of large-scale digital threats. By splitting the release between a public "safe" version and a restricted full version, the company is testing a tiered access model for high-capability AI.
Claude Fable 5 is a Mythos-class model available to the public via Anthropic's API and platform [1], [2]. The company said the model is designed to be safe and cannot be used for cyber-attacks [1], [3]. This specific restriction is intended to address ongoing safety concerns regarding the potential for powerful AI systems to automate malicious hacking or software exploitation [1], [6].
While Fable 5 is accessible to the general public, the full Claude Mythos 5 model remains restricted [2], [4]. Anthropic said it is making the full version available only to trusted organizations [1], [2]. This approach allows the company to maintain oversight of the most potent version of the technology while still providing a high-performance tool to the broader market.
The release occurred in June 2026 [2]. The tiered rollout reflects a growing trend among AI research firms to implement "guardrails" that prevent models from generating harmful code or assisting in offensive cyber operations [3], [5].
Anthropic has not provided specific technical benchmarks for the Fable 5 model in the initial release announcement, but it said the Mythos-class architecture provides a significant leap in reasoning and utility compared to previous iterations [4], [6].
“Claude Fable 5 is a Mythos-class model available to the public via Anthropic's API and platform.”
Anthropic's decision to bifurcate its most powerful model into a public 'safe' version and a private 'full' version highlights the tension between commercial competitiveness and AI safety. By restricting the full Mythos 5 model to trusted partners, the company is attempting to mitigate the risk of catastrophic misuse—specifically cyber-attacks—without conceding the market to competitors who might release less restricted models. This move signals a shift toward a 'governed access' model for frontier AI.
