Open-weights models became the phrase of the week in Washington in late July 2026. Reports suggested some US officials were considering banning US companies from using Chinese open-weights models; many tech companies signed an open letter backing open weights; and Anthropic — the best-known closed-model lab — stood accused of wanting a ban to protect its own business. On July 27, CEO Dario Amodei answered with a post that could hardly be more direct: “Anthropic has never advocated for a ban on open-weights models.”
A Statement Forced by an Open Letter
Amodei opens by clearing the ground: open-weights models without dangerous capabilities are a public good. They cost nothing beyond the compute needed to run them, and they create value for businesses, developers, and researchers. Protectionist bans, he argues, would not address his most serious national-security concerns. The position is not new — he laid out the same argument six months earlier in his essay “The Adolescence of Technology,” and says he has held it consistently for years. The post exists so there is “no doubt.”
Two Nightmare Scenarios
Amodei’s security concerns come in two layers. First, authoritarian governments — not solely the Chinese Communist Party, though he calls the CCP “clearly the most capable threat” — could build models more powerful 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 used by US businesses is irrelevant: the most dangerous model might be one trained in secret and handed only to the People’s Liberation Army for drones and the Ministry of State Security for surveillance. He cites Vice President Vance’s Paris warning and the Intelligence Community’s 2026 Annual Threat Assessment. Second, powerful models could be misused for cyberattacks or biological attacks, or carry serious alignment problems. Here he concedes the open-weights critique has a point: guardrails are hard to apply, usage is hard to monitor, and released weights cannot be withdrawn. But his conclusion cuts the other way — a use ban binds legitimate US businesses, bad actors are unlikely to be legitimate US businesses, and the main thing a ban would protect is US AI companies from competition. “That has never been my goal.”
Three Proposals Instead of a Ban
What Anthropic actually supports is a three-part agenda. One: don’t sell powerful chips or chipmaking equipment to China, and crack down on smuggling and workarounds. By scaling laws, without US chips China cannot build more powerful models than the US — the most direct way to block scenario one. Two: crack down on industrial-scale distillation. Distillation is far more compute-efficient 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 problem for policy interventions, not for open weights as such — many of the companies doing it release open weights, but the state backing matters more than the weights. Three: mandatory pre-release safety testing for all sufficiently capable models, open or closed, covering cyber, biological, and alignment risks. He calls this close to consensus, notes the Trump administration has moved in that direction, and would exempt less capable models from startups and academia entirely. Whether open models are actually riskier, and whether the risk can be mitigated, should emerge from testing rather than be decided in advance — and testing would need to be global to work.
On the Open Letter: Broad Agreement, Two Pushbacks
On the industry letter, Amodei agrees with much of it: open weights expand access to the AI economy, strengthen competition for some use cases, and give customers greater control; distillation concerns belong in targeted legal and commercial frameworks. He rejects two claims — that open-weights models necessarily make safeguards easier to develop, and that broad access to capabilities necessarily helps defenders more than attackers. Biology, he argues, may be strongly attacker-favored: sufficiently capable models could weaponize pandemic-level viruses from widely available materials, while defense is a multi-year operational task at best. Such questions should be answered empirically by rigorous pre-release testing, not assumed in advance.
What It Means for Developers and the Open Ecosystem
Short term, the scare passes: nobody is banning open-weights models as a category, and the dispute over US companies using Chinese open models cools down. Structurally, the battlefield moves to three harder places: 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 more signal worth watching: research from Anthropic and AE Studio on modular training strategies suggests safety features for open-weights models might be built in at training time. The ban debate is over; the testing debate is just starting.
Sources
- Our position on open-weights models — Anthropic
- The Adolescence of Technology — Dario Amodei
- Detecting and preventing distillation attacks — Anthropic
- Modular pretraining research — Anthropic Alignment
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.