Anthropic says it never asked for an open-weights ban, and names what it wants instead
Two industry letters painted Anthropic as the ban lobbyist. Amodei's July 27 answer: no ban, but chip controls, a distillation crackdown, and mandatory testing.
Two letters in one week painted Anthropic as the ban lobbyist. Its July 27 reply names three demands instead.
On July 27, 2026, Anthropic published a position paper stating, in its own words, that "Anthropic has never advocated for a ban on open-weights models." The paper answers a week in which two industry letters, one from nearly 200 startups and one led by Nvidia, treated the company as the loudest voice for exactly that ban. Instead of a ban, it puts three policies on the table.
The two letters it answers
The first letter landed July 22: nearly 200 startups, Y Combinator and Proton among them, wrote to President Trump and Commerce Secretary Howard Lutnick opposing any ban on Chinese open-weight models. We covered it at the time. The second landed July 24, shared by Nvidia's Jensen Huang: roughly 25 companies, including Meta, Microsoft, Mistral, Hugging Face, Dell, and Palantir, signed a letter arguing that open models make AI safer and favor cyber defenders. Three names were conspicuously missing from that list: OpenAI, Google, and Anthropic. The subtext was not subtle. Anthropic, the reading went, wants these models banned.
What the paper actually says
Dario Amodei's answer opens by calling non-dangerous open-weights models "a public good." It then goes after the Huang letter's central claim directly: openness, the paper argues, does not automatically favor defenders over attackers, and a model's open-weight status is "irrelevant" to the threat Anthropic says it worries about most, which is authoritarian states using frontier AI for permanent military advantage. The chokepoint for that threat, in Anthropic's telling, is chips, not licenses.
The three asks
- Export controls on advanced chips and chipmaking equipment going to China, with real enforcement against smuggling. The paper leans on China's limited domestic production capacity as the lever that actually works.
- A crackdown on industrial-scale distillation: harvesting a frontier model's outputs at volume to train a competing model.
- Mandatory pre-release safety testing, covering cyber, bio, and alignment risks, for every sufficiently capable model, open or closed. TechCrunch reports Amodei wants this to grow into a global testing regime with international participation, including China.
The honest tension
The paper does not resolve everything. Mandatory pre-release testing lands differently on an open-weights lab than on a closed one: a closed provider can hold a model behind an API while it patches problems, but an open release is finished the day the weights ship. Startups that cheered last week's letters will notice the third ask regulates them anyway, ban or no ban. And the trigger for this whole cycle, Moonshot's Kimi K3 approaching frontier performance as an open release, is not going away because a position paper published well.
Why a build studio cares
We ship client work on both closed APIs and open models, and this fight decides which of those stay available, at what capability level, and with what compliance overhead attached. When the loudest labs put concrete policy asks in writing, that is the roadmap for what building on AI looks like in two years, so we read the primary documents.
Next step: read Anthropic's position paper and TechCrunch's report on it, and our earlier coverage of the startup letter. If you are weighing open against closed models for a build, write to hello@gattyworks.com.