Amodei: No Open-Weights Ban, But Test Any Capable Model

The interesting move in Anthropic's new position paper, published under Dario Amodei's byline, is what it refuses to defend. Amodei opens by stating flatly that "Anthropic has never advocated for a ban on open-weights models," and frames open-weight models without dangerous capabilities as providing public value to businesses, developers, and researchers. That is not the framing you would expect from a company most of the open-source community assumes has been quietly lobbying against them.

The pivot is where Amodei actually wants the conversation to be. He names two threats, and open-weight releases are only tangential to either. The first is authoritarian governments building more powerful models than the US for military dominance or widespread repression. The second is misuse for cyber and biological attacks, which he argues is harder to mitigate once weights are released, because guardrails cannot be applied at inference time and weights cannot be withdrawn.

From that framing, three concrete asks fall out. The US should not sell powerful chips or chipmaking equipment to China and should crack down on smuggling. Industrial-scale distillation should be treated as a policy problem, because distillation is described as a much more compute-efficient process than training from scratch and could put Chinese work within a few months of the US frontier. And any sufficiently capable model, open or closed, should go through mandatory pre-release safety testing for cyber, biological, and alignment risks.

The honest caveat is that this post arrives right after Anthropic pointedly skipped the Nvidia-led open-weights coalition letter alongside OpenAI, as Axios reported, which is why the framing reads defensive rather than proactive. What the post does not spell out is where the capability threshold for mandatory testing actually sits, who administers it, or what separates industrial-scale distillation from ordinary research use. For smaller open-weight labs the explicit political cover is real. For anyone training at the frontier, the question worth watching is whether that three-part testing regime turns into a bill.