Anthropic’s Claude Fable 5 curbs target China. The backlash came from its own side.

Anthropic built the Claude Fable 5 curbs to keep China’s AI labs out of its most powerful public model. The loudest complaints came from its own side of the firewall.

On 9 June, Anthropic released Claude Fable 5, a tamed, public-facing version of Mythos, the model it had withheld since April over its talent for finding and exploiting security flaws.

Fable 5, the company said, was “safe for general use.” Built into it was a set of classifiers that flag requests touching cybersecurity, biology, chemistry, distillation and frontier model development, then quietly route them to Anthropic’s second-best model, Claude Opus 4.8.

The word that caused the trouble was “quietly.” According to Anthropic’s own system card, the interventions limiting “frontier LLM development” would “not be visible to the user.” AI researchers said they were being silently downgraded, or shadowbanned, the moment Fable 5 decided their machine-learning work looked too much like an attempt to build a rival.

“It felt like Anthropic was saying to the public, we don’t trust anybody else to do AI research,” Will Brown, research lead at Prime Intellect, told Wired. The research firm SemiAnalysis said the model was degrading its GPU-inference work.

The objections came from open-source advocates who usually call Anthropic too closed, and from safety researchers who usually defend it.

Anthropic reversed within two days. “We’re changing Fable 5’s safeguards for frontier LLM development to make them visible,” it said. “We made the wrong trade-off and we apologize for not getting the balance right.” The fix was disclosure, though. The restriction stayed.

What the Claude Fable 5 curbs still do to China

On access, the experts think the restriction works. “Chinese AI developers might find it nearly impossible now to use Anthropic’s latest model to accelerate their own model development,” Kyle Chan, a fellow at the Brookings Institution, told the South China Morning Post. Claude has never been officially sold in China, but developers there long relied on workarounds.

Fable 5’s classifiers are much harder to route around.

Anthropic says those workarounds were industrial. In February it accused DeepSeek, Moonshot and MiniMax of running more than 16 million exchanges through roughly 24,000 fraudulent accounts to distil Claude’s coding and reasoning. The White House took up the theme in April, calling the campaigns industrial-scale theft.

The dependency runs both ways

Then the picture inverts. US firms are increasingly reaching for China’s DeepSeek because it is cheap, while Chinese developers keep buying grey-market Claude through proxy “transfer stations” that resell it on Taobao and Telegram at as little as a tenth of list price, according to the Oxford China Policy Lab.

Beijing is busy building walls of its own: this week Meta began dismantling its $2 billion purchase of Manus, the Chinese-founded, Singapore-based startup, after regulators ordered the closed deal reversed.

Critics see less principle than protectionism, noting that Anthropic trained Claude on scraped, copyright-protected text before it began policing anyone else’s extraction. The curbs also land as Anthropic heads for the public market, having filed confidentially for an IPO at a $965bn valuation.

A model too open invites distillation; a model too closed invites the backlash Fable 5 just triggered. Anthropic has drawn its line. Whether classifiers can hold it, when the first version lasted barely two days, is the question its rivals on both sides of the Pacific are already testing.