Anthropic restricted access to the most capable features of Claude Fable 5, the public version of its Mythos-class model, embedding classifiers that detect and downgrade queries related to cybersecurity, biology, chemistry and frontier model development.
When the system flags a request, it automatically routes the response to Claude Opus 4.8, an older, less capable model. The user is now told this is happening, following a backlash over the original policy of silent degradation.
The safety rationale is genuine. Mythos alarmed the industry during internal testing with its ability to identify unknown cybersecurity vulnerabilities in production software. Restricting that capability in a public release is defensible.
But the restrictions go beyond safety. They include anti-distillation measures, which prevent other AI developers from using Fable 5's outputs to train their own models. Experts say this layer is aimed directly at Chinese AI labs.
New containment
Winston Ma at New York University described the shift as a move from hardware containment to algorithmic containment. For three years, the US strategy for limiting Chinese AI capability focused on export controls on chips. Deny the hardware, deny the capability.
That approach has been partially circumvented. Chinese companies have acquired chips through third countries, developed domestic alternatives through SMIC, and found workarounds that the Commerce Department has struggled to close.
Anthropic's approach embeds the containment in the model itself. The restriction travels with the product. It does not matter where the user is or what hardware they are using. If the model detects an attempt to extract frontier capability, it downgrades the response.
Impact on Chinese labs
Kyle Chan of the Brookings Institution said Chinese developers may now find it nearly impossible to use Anthropic's latest model to accelerate their own development. The anti-distillation measures close a pathway that smaller labs have relied on: using outputs from a frontier model to train a cheaper, local alternative.
Jack Jiang Zhenhui at the University of Hong Kong Business School said the controls raise the barrier to entry for smaller teams that depend on low-cost catch-up strategies. The labs most affected are not the large, state-backed operations with their own research capacity but the mid-tier developers that relied on Western models as a shortcut.
Ronald Sun of IntelliGen AI warned that Fable 5's strengths in scientific research, particularly drug design, are now effectively isolated from an industry that is routed to an older model the moment it approaches frontier-level work.
Safety and strategy are not mutually exclusive
Anthropic framed the restrictions around safety: "You should have visibility into the safeguards we have in place, and why." The company said it will reduce unnecessary obstructions while maintaining limits on frontier development queries.
The safety argument and the competitive argument point in the same direction. Preventing distillation protects against misuse. It also protects Anthropic's commercial position by ensuring that the capability gap between its models and its competitors' models cannot be closed by using Fable 5 as a training resource.
Experts were direct about this. Academics and industry observers described the anti-distillation measures as serving both safety and moat-building purposes. The restrictions protect users from dangerous capabilities. They also protect Anthropic from competition.
The chip export controls tried to contain Chinese AI development at the hardware layer. Anthropic is now containing it at the software layer. The containment has moved from the factory to the model, and the model goes everywhere the internet reaches.
