Anthropic wants tests, not bans, as OpenAI and Google back open weights

Dario Amodei spent part of this week defending his company from a charge nobody actually made against it. After a string of criticism, the Anthropic CEO published a letter insisting, in plain terms, that Anthropic has never called for open-weight models to be banned. He calls models without dangerous capabilities a public good, something that costs nothing beyond the compute needed to run them. What he wants instead is a three-part policy: tighter export controls on advanced chips and chipmaking equipment, a crackdown on industrial-scale model distillation, and mandatory safety evaluations for any model deemed sufficiently capable before it ships.

That last phrase is doing a lot of work. Amodei frames the testing idea as close to consensus, pointing out that the Trump administration has been moving in that direction and that recent industry proposals would test models regardless of license or origin, while carving out an exemption for less capable models from startups and academia. Sounds reasonable until you ask where the exemption line sits. Nobody in the letter says. And once some models are exempt and others aren't, you've built a capability threshold in everything but name.

The letter never names who would set that threshold, who would run the tests, or which metric would decide capability. Benchmark scores, training compute, risk-category evaluations — take your pick, because whoever picks determines which models get delayed and which sail through. That's not a small detail. It's release timing for every lab that publishes weights, decided by a number that doesn't exist yet.

Meanwhile the coalition lines are hardening. The Open Weights and American AI Leadership letter has picked up 50 signatories, including recent additions OpenAI and Google alongside Microsoft, Nvidia, Meta, AMD, Cisco, Cloudflare, GitHub and Ollama. Anthropic and Amazon are conspicuously absent. Nvidia went further this week, launching its own Open Secure AI Alliance with more than 35 founding members — Microsoft, IBM, Hugging Face, Palantir, Red Hat, CrowdStrike, the Linux Foundation — arguing that defenders need open frontier models to do forensic work, citing a recent Hugging Face incident where closed tooling couldn't help and an open-weight model did the job instead. Anthropic isn't in that group either.

Amodei does address the obvious jab: that a domestic ban would mainly protect U.S. AI companies from competition rather than stop bad actors, and he says that's never been his goal. Fair enough, though it's an easy thing to say when your own company doesn't release open-weight models in the first place. He also concedes that his testing regime would need to be global, meaning even the CCP would need to sign on, and he thinks that might be achievable. Export controls and anti-distillation rules are things Washington can do unilaterally. A worldwide capability-testing regime is an entirely different lift, and it's the piece of this plan that actually touches open-weight models — resting on international agreements that, right now, simply don't exist.