Anthropic says it does not want open-weight models banned

Anthropic did not sign the industry letter defending open-weight models, but that does not mean it is asking for them to be banned. On July 27, Dario Amodei issued a direct clarification: the company has never advocated a categorical ban on open-weight models and regards models without dangerous capabilities as a public good.

The clarification matters because the debate had collapsed two different positions into a simple “open versus closed” divide. Anthropic’s actual position is narrower. It accepts the benefits of access, competition and user control, while arguing that highly capable systems should pass safety tests even when their weights can be downloaded.

The letter Anthropic did not sign

The letter is titled “Open Weights and American AI Leadership” and was published on July 24. It asks Washington to avoid premature restrictions on downloadable models and argues that openness can strengthen competition, technological sovereignty and defensive capabilities.

The signatory list expanded after publication. Microsoft, Nvidia, Meta, IBM, Hugging Face and dozens of other organizations are listed; Google and OpenAI joined over the weekend. Anthropic was still absent when Amodei published his response.

The three measures Anthropic does support

• Keep advanced chips and chipmaking equipment out of the hands of authoritarian governments, while also targeting smuggling.
• Act against industrial-scale distillation operations that use the outputs of more advanced systems to train other models.
• Require sufficiently capable models, open or closed, to undergo testing for cyber, biological and alignment risks before release.

Amodei also separates a model’s country of origin from its distribution model. In his view, preventing a legitimate US company from using a Chinese model would not stop a malicious actor from running it beyond US jurisdiction. He therefore wants policy aimed at capability and strategic chokepoints, not at a commercial label.

Why it matters

If this distinction shapes regulation, a startup or academic lab could publish a small model without facing the same regime as a frontier system. At the same time, any open or closed model that reaches sensitive capability thresholds would need to provide safety evidence before release.

The remaining disagreement is empirical. Nvidia and the letter’s signatories argue that open access improves defence by enabling inspection and adaptation. Anthropic says the balance may favour attackers in areas such as biology. Both claims require evidence rather than slogans.

What is known — and what is not

Amodei’s words, Anthropic’s three preferred measures and the company’s absence from the letter are all confirmed. Anthropic has not, however, published a complete legislative design explaining who would set the thresholds, how global tests would work or what would happen if a country refused to participate.