Update: Amodei says Anthropic never backed an open-weights ban
After sitting out the industry's open-weights letter, the CEO says his real asks are China chip curbs, a distillation crackdown, and mandatory safety tests.
Copy markdown
The record, set straight
Dario Amodei: "Anthropic has never advocated for a ban on open-weights models," and he doesn't view bans as "a useful measure." It's a direct answer to the Nvidia-Meta-Microsoft-Hugging Face-Mistral letter that Anthropic and OpenAI declined to sign, recasting a silence many builders read as opposition.
What he wants instead of a ban
Three levers, not one: keep U.S. curbs on China's access to advanced chips, crack down on distillation of frontier models, and stand up a mandatory model-safety-testing organization, ideally global and applied to China too. For builders that reads as a compliance layer, not a takedown of open weights.
The real target isn't the model on your disk
Amodei frames the threat as authoritarian governments, with the CCP "most capable" among them, chasing "permanent military superiority" and repression, not the open checkpoints you run locally. He calls non-dangerous open weights beneficial, costing "nothing besides the compute needed to run them."
Why it matters for your stack
This fight decides whether the cheap, capable Chinese models builders now lean on, Kimi K3, GLM-5.2, DeepSeek V4, Qwen, stay usable in the U.S. Amodei's chips-plus-testing path is gentler on the weights themselves than the separate Washington ban push, but a mandatory-testing regime could still gate which open models ever ship.
What to do this week
Nothing is banned today, no rule is law. But if a product depends on a Chinese open model, wire in a Western-weights fallback (Llama, Mistral, Gemma) and pin versions you can self-host. The policy path now runs through testing and paperwork, which lands as gating long before it touches your inference calls.