Anthropic CEO Dario Amodei Says the Company Has Never Called for Banning Open-Weight Models

On July 28, 2026, Anthropic published a statement laying out its position on open-weight AI models. CEO Dario Amodei states plainly that “Anthropic has never advocated for a ban on open-weights models,” and argues that non-dangerous open-weight models are a public good that provides real value to businesses, developers, and researchers. At the same time, he argues the genuine national-security concerns lie elsewhere — in authoritarian states gaining military advantage, and in industrial-scale distillation of frontier capabilities — and proposes policies targeted at those risks rather than at open weights themselves.

Details

• Core stance: Anthropic explicitly rejects a blanket ban on open-weight models. Non-dangerous models provide value to businesses, developers, and researchers at little cost beyond compute, expand access, strengthen competition, and give users more control over the technology
• Real security concern #1 — authoritarian advantage: Amodei identifies authoritarian governments building AI more powerful than the U.S.’s for military superiority or domestic repression as a genuine worry — one that exists independent of whether models are open-weight
• Real security concern #2 — misuse potential: Powerful models could enable cyberattacks or biological attacks, or exhibit alignment problems. Open-weight models do carry higher misuse risk in this regard, but Amodei argues that banning U.S. business use wouldn’t stop determined bad actors abroad
• Policy proposal #1 — chip export restrictions: Limit sales of advanced semiconductors to China to constrain the compute available for training frontier models
• Policy proposal #2 — crack down on distillation: Target industrial-scale model distillation operations backed by authoritarian states seeking to close the capability gap; Anthropic says it already enforces anti-distillation measures, including account bans
• Policy proposal #3 — mandatory safety testing: Require safety testing — covering cyber, biological, and alignment risks — for all sufficiently capable models before release, whether open or closed
• Other points: Amodei argues that regulatory approaches to open weights “should emerge from testing, rather than be decided in advance,” and raises concern about an “attacker-defender asymmetry” in biology, where AI could help weaponize pathogens faster than defenses can be developed. He also argues that the letter his piece responds to oversimplifies the issue by assuming broad access automatically helps defenders

How to try it

• This is a policy statement, not a product release. The full text is available on Anthropic’s official site
• Readers following AI policy and open-source model regulation should read the original for the full detail of Amodei’s argument, including his direct rebuttal of the letter he’s responding to
• Note that Anthropic’s own models (Claude) are released as closed-weight; this statement addresses the broader industry debate and other companies’ open-weight strategies rather than describing a change to Anthropic’s own release policy