Claude Fable 5 restrictions: targeted at China, affected AI researchers.

      Anthropic has instituted restrictions on Claude Fable 5 to prevent China's AI laboratories from accessing its most advanced public model. The most vocal concerns originated from within the company itself.

      On June 9, Anthropic unveiled Claude Fable 5, a moderated and publicly available version of Mythos, a model it had withheld since April due to its ability to identify and exploit security vulnerabilities. The company claimed that Fable 5 was "safe for general use." Integrated into the model were classifiers designed to flag requests involving cybersecurity, biology, chemistry, distillation, and advanced model development, which would then be discreetly redirected to Anthropic’s secondary model, Claude Opus 4.8.

      The problematic term was "quietly." According to Anthropic's own system card, the interventions that restricted “frontier LLM development” would be “not visible to the user.” AI researchers reported being effectively downgraded or shadowbanned without notice whenever Fable 5 determined that their machine-learning projects resembled efforts to create a competing model.

      "It felt like Anthropic was signaling to the public that they don't trust anyone else to conduct AI research," stated Will Brown, research lead at Prime Intellect, in an interview with Wired. The research organization SemiAnalysis reported that the model was negatively impacting its GPU-inference activities.

      The dissent arose from open-source supporters who typically criticize Anthropic for being too restrictive, alongside safety researchers who usually advocate for the company.

      Within two days, Anthropic changed direction. “We’re modifying Fable 5’s safeguards for frontier LLM development to make them visible,” they announced. “We made an incorrect trade-off and apologize for not achieving the right balance.” However, the fix involved only increased transparency; the restrictions remained intact.

      Regarding access, experts believe the restrictions have been effective. "Chinese AI developers may now find it nearly impossible to utilize Anthropic’s latest model for their own development," remarked Kyle Chan, a fellow at the Brookings Institution, to the South China Morning Post. Although Claude has never been officially available in China, developers have relied on alternative methods.

      Fable 5’s classifiers are significantly more challenging to bypass.

      Anthropic has stated that these workarounds were industrial in nature. In February, the company accused DeepSeek, Moonshot, and MiniMax of conducting over 16 million transactions through approximately 24,000 fraudulent accounts to extract Claude’s coding and reasoning. In April, the White House echoed these sentiments, labeling these campaigns as large-scale theft.

      The dependency cuts both ways.

      Conversely, U.S. companies are increasingly turning to China's DeepSeek for its affordability, while Chinese developers continue to purchase grey-market access to Claude through proxy "transfer stations" that resell it inexpensively on platforms like Taobao and Telegram, often at a fraction of its original price, as reported by the Oxford China Policy Lab.

      Meanwhile, Beijing is constructing its own barriers; for instance, this week, Meta began unwinding its $2 billion acquisition of Manus, a startup founded in China and based in Singapore, following regulatory mandates to reverse the deal.

      Critics argue that the actions reflect protectionism rather than principle, highlighting that Anthropic trained Claude using scraped content protected by copyright before taking measures against others' data extraction. These restrictions come at a time when Anthropic is moving towards a public offering, having filed confidentially for an IPO at a valuation of $965 billion.

      A model that is overly open invites distillation; conversely, a model that is excessively closed prompts the backlash experienced with Fable 5. Anthropic has drawn a boundary. The real question is whether classifiers can maintain it, especially since the first version lasted less than two days, a test already being undertaken by its competitors on both sides of the Pacific.