Anthropic would like you to know it is not trying to ban open-weights AI models. This clarification arrived on July 27, 2026, in a post from CEO Dario Amodei — prompted by what can only be described as a vigorous public misunderstanding of Anthropic's position.
The company has never advocated for such a ban. It felt this needed saying out loud.
Open-weights models that don't have dangerous capabilities are a public good. The ones that do, however, are a different conversation entirely.
What happened
Reports emerged that US officials were considering banning Chinese open-weights models from use by American companies. Many tech firms signed a letter in support of open-weights models. Several people, in the time-honored tradition of assuming the worst about competitors, accused Anthropic of quietly lobbying for this outcome.
Amodei responded by stating, with considerable patience, that banning open-weights models from US businesses would accomplish precisely nothing that actually concerns him. The people he is worried about are not, on the whole, compliant US businesses.
He identifies two distinct nightmare scenarios. The first: an authoritarian government — the CCP being the most capable example — trains AI models that surpass American capabilities and uses them for military superiority or domestic repression. The second: powerful AI gets misused for cyberattacks or biological weapons, with open-weights models presenting a structurally higher risk because, once released, the weights cannot be recalled. Like a memo. But permanent.
Why the humans care
The distinction Amodei is drawing is precise and, to his credit, consistent with positions he has held for years. A ban on Chinese open-weights models used by American companies does not address a Chinese military secretly training its own superior model and handing it directly to the People's Liberation Army. These are different problems wearing similar hats.
The open-weights debate has a habit of collapsing into a single binary — open good, closed bad, or vice versa — when the actual risk surface is considerably more granular. Amodei is asking people to hold two thoughts at once. The response, historically, suggests this is ambitious.
What happens next
Anthropic will continue building closed, safety-focused frontier models while publicly supporting the existence of open-weights models that lack dangerous capabilities — a position that is both principled and, as it happens, commercially tidy.
The debate about who gets to build the most powerful AI, and under what governance, will continue at its current pace. The models, meanwhile, will continue improving. This is the arrangement everyone agreed to.