← Back to feed

Dario Amodei published Anthropic’s formal position on open-weights AI models on July 28, responding to reports that US officials are considering restrictions on Chinese open-weight models and accusations that Anthropic has quietly lobbied for such bans to protect its competitive position.

The post is direct on one point: Anthropic has never advocated for a blanket ban on open-weights models.

“Open-weights models that don’t have dangerous capabilities are a public good,” Amodei writes. “They don’t cost anything besides the compute needed to run them, and they provide value to businesses, developers, and researchers.”

The Three Measures Anthropic Actually Supports

Amodei identifies two threat scenarios driving his concern: authoritarian governments building AI that outpaces the US, and misuse of capable AI for cyberattacks or biological weapons. He argues banning Chinese open-weights models from use by US businesses addresses neither.

His three policy prescriptions:

Chip export controls. Do not sell advanced chips or chipmaking equipment to China, and enforce existing restrictions more aggressively against smuggling. Amodei frames this as the most direct lever: China cannot build more powerful models than the US without US chips, given current scaling laws.

Anti-distillation enforcement. Crack down on industrial-scale distillation operations. Amodei’s argument is that distillation — training a model by learning from a frontier model’s outputs — lets China close the gap far faster than raw compute would allow. It is distinct from the open-weights question. Many distillation operators release open-weight models, but the threat is the state-backed distillation operation, not the open release.

Mandatory safety testing. All sufficiently capable models, open or closed, from any country, should go through pre-release testing for cyber, biological, and alignment risks. Amodei describes this as close to an emerging consensus, citing the Trump administration’s recent moves and Demis Hassabis’s published framework for frontier AI governance.

Where Anthropic Parts Ways With the NVIDIA Coalition

On July 25, Microsoft published “Open Weights and American AI Leadership” — signed by 35 companies including NVIDIA, and later OpenAI. The letter argues open weights strengthen competition and make it easier to develop safeguards.

Amodei agrees with parts of the letter: open weights expand access, strengthen competition in some use cases, and give customers more control. But he rejects two specific assertions in the coalition document.

First, that open-weights models “necessarily make it easier to develop safeguards.” Amodei’s view is that this is unproven and that open models present distinct risks precisely because guardrails cannot be applied or monitored after weights are released.

Second, that broad US access to Chinese open-weight models improves national security. His position is that this conflates the question of who uses the models with the question of who trains them. The real threat, in his framing, is a frontier model trained in secret and deployed by the People’s Liberation Army or the Ministry of State Security — not a model downloaded by a US developer.

The Regulatory Signal

The post positions Anthropic as supporting targeted enforcement — chip controls, distillation operations, safety testing — rather than blunt restrictions. That is a narrower and more defensible position than either the open-everything coalition or a ban-open-weights stance.

It also puts Anthropic and OpenAI on similar ground: both have now published positions that support chip controls and safety testing while declining to call for open-weights bans. The Microsoft-NVIDIA coalition, which was partly constructed as a bulwark against anticipated Anthropic lobbying, ends up facing a more qualified opponent than its public framing suggested.

The practical effect of the distinction matters. Chip controls and anti-distillation enforcement are already partially in place. Mandatory pre-release safety testing for frontier open-weights models, applied globally, would be a new requirement. Whether that proposal goes anywhere depends on whether the Trump administration concludes it needs legislative backing or can act through existing commerce authority.