On June 9, 2026, Anthropic released Claude Fable 5, a publicly available version of its internal Mythos model that had been restricted since April due to security concerns. The model incorporates automated classifiers designed to detect queries related to cybersecurity, biology, chemistry, and frontier model development. When triggered, these classifiers silently reroute users to Claude Opus 4.8. Critics, including AI researchers and open-source advocates, labeled this silent redirection a 'shadowban' on machine-learning development work, leading Anthropic to announce a policy reversal two days later. The company now makes these safeguard interventions visible to users, though the underlying restrictions remain in effect.

Anthropic implemented these controls primarily to prevent developers in China from leveraging its most powerful model to accelerate their own AI research. Although Claude is not officially sold in China, researchers note that developers have historically bypassed restrictions using proxy services. According to Anthropic, firms such as DeepSeek, Moonshot, and MiniMax previously utilized approximately 24,000 fraudulent accounts to process over 16 million exchanges in efforts to distill Claude’s capabilities. In April, the White House characterized these activities as industrial-scale theft. Experts such as Kyle Chan of the Brookings Institution suggest that the new, more robust classifiers may now make it nearly impossible for Chinese developers to use Claude for model development workarounds.

The restriction arrives amid increased scrutiny of cross-border AI access. While US firms have increasingly utilized cheaper models from DeepSeek, Chinese developers continue to access grey-market Claude models via third-party transfer stations on platforms like Taobao and Telegram, often at one-tenth of the list price. Meanwhile, Meta recently halted a $2 billion acquisition of the Chinese-founded startup Manus following regulatory intervention. Anthropic faces criticism regarding its own data practices, with observers pointing out that it trained its models on scraped, copyrighted text before imposing its own extraction policies. The company, which recently filed confidentially for an IPO at a $965bn valuation, maintains that these curbs are necessary to secure its technology while navigating a complex landscape of global competition and regulatory pressure.

On June 9, 2026, Anthropic released Claude Fable 5, a publicly available version of its internal Mythos model that had been restricted since April due to security concerns. The model incorporates automated classifiers designed to detect queries related to cybersecurity, biology, chemistry, and frontier model development. When triggered, these classifiers silently reroute users to Claude Opus 4.8. Critics, including AI researchers and open-source advocates, labeled this silent redirection a 'shadowban' on machine-learning development work, leading Anthropic to announce a policy reversal two days later. The company now makes these safeguard interventions visible to users, though the underlying restrictions remain in effect.

Anthropic implemented these controls primarily to prevent developers in China from leveraging its most powerful model to accelerate their own AI research. Although Claude is not officially sold in China, researchers note that developers have historically bypassed restrictions using proxy services. According to Anthropic, firms such as DeepSeek, Moonshot, and MiniMax previously utilized approximately 24,000 fraudulent accounts to process over 16 million exchanges in efforts to distill Claude’s capabilities. In April, the White House characterized these activities as industrial-scale theft. Experts such as Kyle Chan of the Brookings Institution suggest that the new, more robust classifiers may now make it nearly impossible for Chinese developers to use Claude for model development workarounds.

The restriction arrives amid increased scrutiny of cross-border AI access. While US firms have increasingly utilized cheaper models from DeepSeek, Chinese developers continue to access grey-market Claude models via third-party transfer stations on platforms like Taobao and Telegram, often at one-tenth of the list price. Meanwhile, Meta recently halted a $2 billion acquisition of the Chinese-founded startup Manus following regulatory intervention. Anthropic faces criticism regarding its own data practices, with observers pointing out that it trained its models on scraped, copyrighted text before imposing its own extraction policies. The company, which recently filed confidentially for an IPO at a $965bn valuation, maintains that these curbs are necessary to secure its technology while navigating a complex landscape of global competition and regulatory pressure.