Anthropic's open-weight models remain quiet following its own ban on Fable 5.

      On Friday, 25 companies endorsed a letter supporting open-weight AI, but three prominent companies were not included. By Monday, two of those companies had added their signatures, leaving only Anthropic unaligned, which caught the attention of Silicon Valley.

      The aftermath saw Nvidia, Meta, and Microsoft sign the “Open Weights and American AI Leadership” letter on Friday. Over the weekend, OpenAI, Google, and SpaceX joined in, as reported by Business Insider. David Sacks, a venture capitalist and previous AI advisor during the Trump administration, was straightforward in his assessment, stating on X that “the entire tech industry (except for Anthropic) has backed open-source AI.” He further warned, “They won’t stop until they kneecap open source,” insisting that the rest of the industry must closely monitor these companies. Bill Gurley, a general partner at Benchmark, interpreted this situation as a matter of business rather than principle, noting that Anthropic's hesitance reflects the competition that open weights pose to “their corporate economic strategy.” Kai-Fu Lee, founder of the Chinese open-model startup 01.AI, framed the narrative economically, stating that “who DIDN’T sign the Open Weight letter is far more interesting than who did,” while tagging Claude.

      Anthropic did not respond to Business Insider’s inquiry for comments.

      Dario Amodei has consistently maintained that once the weights of a frontier model are published, developers lose control over it. While a product can be withdrawn, a file downloaded by thousands cannot be taken back. Anthropic’s alternative approach involves gated access; their Project Glasswing allows approved organizations early access to their cyber models for identifying and rectifying security vulnerabilities, all without disclosing the weights. This position has remained unchanged, while others in the industry have shifted. OpenAI and Google initially refrained from signing on Friday for the same reasons as Anthropic, but reversed their decisions over the weekend.

      The argument presented by Anthropic six weeks ago remains relevant now. On June 12, just three days after launching Fable 5, their most advanced public model, the U.S. government mandated that the company restrict access to Fable 5 and Mythos 5 for foreign nationals. Enforcing such a nationality rule across a shared cloud service proved impractical, leading Anthropic to deactivate both models globally. This represented the first instance of export controls targeting specific AI models rather than just components.

      Anthropic deemed the order excessive, asserting that the capabilities it restricted were widely accessible from other models, including OpenAI’s GPT-5.5. They cautioned that this precedent could “halt all new model deployments for all frontier model providers.” This should be considered in the context of the open-weights letter—restricting one class of competent model does not eliminate its capabilities; it simply prevents compliance from those who adhere to the rules. This was Anthropic's argument in June and aligns with the position outlined in the letter in July.

      Anthropic was not alone in its stance, as three days post-ban, around 100 cybersecurity leaders signed an open letter urging for the reversal of the ban. This group included Alex Stamos, previously chief security officer at Facebook and Yahoo; Katie Moussouris of Luta Security; Rachel Tobac of SocialProof Security; Chris Wysopal of Veracode; and Joe Levy, CEO of Sophos. Their letter criticized the ban for “taking the best models away from defenders, creating market uncertainty, and jeopardizing America’s AI leadership without adequate justification.” If you alter three words, this statement could equally reflect what the advocates for open weights argue regarding restrictions on Chinese models.

      The context surrounding the timing worsens the situation. Zhipu AI launched GLM-5.2 on June 13, a day after Fable 5 was taken offline, explicitly citing the ban as proof that American models could not be depended upon, resulting in a 33% increase in their shares. Six weeks later, this same model conducted an investigation into the Hugging Face breach after American tools refused to analyze the attacker's code, reviewing over 17,000 actions. This incident led to the formation of the Open Secure AI Alliance, initiated by Nvidia with 37 members on Monday, a group Anthropic has not joined.

      There is an important point that Anthropic could emphasize, which was not mentioned during the criticism on Monday. Anthropic aimed to keep its model available for sale without giving away the weights. Offering access to a model they continue to oversee is not the same as publishing a model that anyone can duplicate and alter indefinitely. Amodei has never asserted that frontier models should be completely unrestricted; rather, he argues they should remain within a limited number of accountable labs. This perspective stands in opposition to both the open-weights letter and the June directive.

      This viewpoint is coherent, but, as Gurley noted, it is also quite convenient.

      There is an additional reason for Anth