Summary: In late July 2026, Dario Amodei said Anthropic never advocated a ban on open-weights models, calling models without dangerous capabilities a public good. He instead supported chip export controls, action against industrial-scale distillation, and mandatory pre-release safety testing for sufficiently capable open or closed models. He argued testing should determine risk rather than treating openness itself as the decisive issue.
In late July 2026, open-weights models turned into Washington’s phrase of the week. Some US officials were reported to be weighing a ban on American companies’ use of Chinese open-weights models; a long list of tech companies then signed an open letter in support of open weights; and Anthropic — the most visible lab in the closed-model camp — found itself accused of backing a ban to shield its own business. On July 27, CEO Dario Amodei answered in a post of his own and left no room for interpretation: “Anthropic has never advocated for a ban on open-weights models.”
A Statement the Open Letter Forced
Amodei begins by clearing the ground: open-weights models that lack dangerous capabilities are a public good. Beyond the compute required to run them, they cost essentially nothing, and they deliver value to businesses, developers, and researchers. A protectionist ban, he argues, would do nothing about his most serious national-security worries. The argument itself is not new — he set it out six months earlier in his essay “The Adolescence of Technology,” and says his view has held consistent for years. The post exists so that there is “no doubt.”
Two Nightmare Scenarios
His security worries arrive in two layers. The first: authoritarian governments — not China’s alone, though he calls the CCP “clearly the most capable threat” — could build models stronger than America’s and convert that lead into permanent military superiority or deep repression of their own people. Whether such models are open-weights, or whether US businesses run them, is beside the point: the most dangerous model of all might be one trained in secret and handed only to the People’s Liberation Army for drones and to the Ministry of State Security for surveillance. He points to Vice President Vance’s warning in Paris and to the Intelligence Community’s 2026 Annual Threat Assessment. The second layer: powerful models could be turned to cyberattacks or biological attacks, or could carry serious alignment failures. Here he grants the open-weights critique some ground — guardrails are hard to attach, usage is hard to monitor, and released weights cannot be called back. His conclusion, though, runs the other way. A use ban would bind only legitimate US businesses; bad actors are unlikely to be legitimate US businesses; and what such a ban would really protect is US AI companies from competition. “That has never been my goal.”
Three Proposals Instead of a Ban
What Anthropic does back is a three-part agenda. First: stop selling powerful chips and chipmaking equipment to China, and go after smuggling and workarounds. By scaling laws, China cannot build models stronger than America’s without US chips — the most direct way to block the first scenario. Second: go after industrial-scale distillation. Distilling is far cheaper in compute than training from scratch; it lets China partly evade chip restrictions and could pull the Chinese frontier to within months of the US frontier. That is a job for policy intervention — many of the firms doing it do release open weights, but he treats the state backing as the real issue, not the weights themselves. Third: mandatory pre-release safety testing for every sufficiently capable model, open or closed, spanning cyber, biological, and alignment risks. He describes this as close to consensus, notes that the Trump administration has moved in that direction, and would exempt less capable models from startups and academia entirely. Whether open models are in fact riskier, and whether the risk can be mitigated, is a question for test results rather than prior assumption — and testing would have to be global to mean anything.
On the Open Letter: Broad Agreement, Two Pushbacks
Much of the industry letter Amodei agrees with: open weights widen participation in the AI economy, sharpen competition in some use cases, and hand customers greater control; distillation concerns belong in targeted legal and commercial frameworks. Two of its claims he will not accept — that open-weights models necessarily make safeguards easier to develop, and that broad access to capabilities necessarily helps defenders more than attackers. The outcome, he argues, may well be the reverse, and biology is his example: a capable enough model might weaponize pandemic-level viruses from widely available materials, while defense is at best a multi-year operational task. Questions like that should be settled empirically, by rigorous pre-release testing, not assumed in advance.
What It Means for Developers and the Open Ecosystem
In the short term the scare passes: open-weights models are not being banned as a category, and the dispute over US companies using Chinese open models cools down. Structurally, the battlefield shifts to three harder fronts: enforcement of chip export controls, the legal treatment of distillation, and the design of mandatory testing thresholds — where the threshold lands will decide which open models can legally exist. One further signal to watch: modular training research published by Anthropic with AE Studio suggests safety features for open-weights models might be built in at training time. The ban debate is finished; the testing debate has barely begun.