After OpenAI, Google, and SpaceX all signed an open-source coalition letter, Anthropic remains the only major frontier AI lab that hasn't put its name on it. Now, CEO Dario Amodei is finally breaking the silence.
Amodei says the company never called for a total ban on open-weight models. He acknowledges that open-weight models can lower costs, increase competition, and let customers deploy models themselves. Models without dangerous capabilities, he says, are a "public good."
But that doesn't mean he thinks open models are inherently safer, or that defenders necessarily benefit more than attackers. Once weights are released, safety restrictions can be removed—and there's no taking them back.
So what does Anthropic want? Not a ban, but tighter controls on chips and model distillation—the process of using a large model's outputs to train a smaller one. That's where the real risk lies, in their view.