July 28, 2026. Anthropic CEO Dario Amodei published a post yesterday titled Our position on open-weights models, opening with an unusually blunt line: Anthropic has never advocated for a ban on open weights models. It is a direct response to a policy fight that broke open last week, and the reason it matters to a business owner rather than a policy analyst is that the models under discussion are already inside ordinary go to market tools.
How the week unfolded
- Axios reported on July 20 that some US officials were weighing restrictions on US companies using Chinese open weights models.
- On July 24, a three page letter titled Open Weights and American AI Leadership was published and hosted by Nvidia, and Jensen Huang used his first post on X to promote it. It asks Washington to avoid what it calls premature restrictions on downloadable AI models. Reporting from CNBC and Tom's Hardware put the launch list at 25 companies including Nvidia, Microsoft, Meta, IBM, Dell, Palantir and Hugging Face, roughly doubling to about 50 within a day. OpenAI, Anthropic and Google were absent at launch; OpenAI and Google were reported to have signed later, alongside AMD, Cisco, Cohere, GitHub, Cloudflare and Block. Anthropic has not signed.
- On July 27, Amodei answered. He calls open weights models without dangerous capabilities a public good, says protectionist bans would not address his actual national security concerns, and states plainly that shielding US AI companies from competition has never been his goal.
- What Anthropic does support is narrower and specific: keep powerful chips and chipmaking equipment out of China and crack down on smuggling, deter industrial scale distillation, and require pre release safety testing of all sufficiently capable models, open and closed alike.
- He does not concede the whole argument. Amodei says open weights carry potentially higher risk because guardrails are hard to apply and, once released, weights cannot be withdrawn, and he rejects the letter's assertion that broad access necessarily helps defenders more than attackers. On distillation, he says it can bring the Chinese frontier to within a few months of the US frontier.
Why this is not an abstract debate
Here is the part most coverage skipped. Chinese open weight models are not a hypothetical import waiting at the border. They are already shipping inside mainstream sales tooling. On July 16 the GTM platform Clay added Kimi K2.6 and GLM 5.2 as selectable models inside Claygent, its research agent, explicitly to run research at lower cost. If your team picked the cheap option in a dropdown, a Chinese open weight model may already be reading your prospect list.
Meanwhile the model everyone argued about all weekend still is not downloadable. Moonshot promised Kimi K3 open weights for July 27. Checking the Hugging Face model API for the moonshotai organisation this morning returned no K3 repository, a day past the stated deadline. The policy fight is running ahead of the artefact.
What it means for operators
Nothing here requires you to rip anything out today. No restriction exists yet, and the companies arguing about it disagree in good faith about a genuinely hard question. What it does require is that you know the answer to one question you probably cannot answer right now: which model is running each step of your stack.
Three things worth doing this week. Open each AI tool you pay for and write down the model actually selected, including the defaults you never changed. Where a tool lets you choose, note whether the cheap option is an open weight model and where it is hosted, because self hosted weights and a hosted API from a sanctionable vendor are two very different exposures. And make sure every automation that calls a model has a tested fallback, so a vendor outage or a policy change is a configuration change rather than a rebuild. That is the same portability principle that protects a lead generation system from any single supplier, and it is the first thing we design for when we build an AI automation stack or place an AI engineer on a team.
Frequently Asked Questions
In a July 27, 2026 post, CEO Dario Amodei stated that Anthropic has never advocated for a ban on open weights models and called open models without dangerous capabilities a public good. He supports three measures instead: restricting powerful chips and chipmaking equipment to China plus a smuggling crackdown, deterring industrial scale distillation, and mandatory pre release safety testing for all sufficiently capable models, open and closed.
A three page letter titled Open Weights and American AI Leadership, published and hosted by Nvidia on July 24, 2026 and promoted by Jensen Huang in his first post on X. It asks Washington to avoid premature restrictions on downloadable AI models. Reporting put the launch list at 25 companies including Nvidia, Microsoft, Meta, IBM, Dell, Palantir and Hugging Face, roughly doubling to about 50 signatories within a day. Anthropic did not sign.
Yes, in some cases. The GTM platform Clay added Kimi K2.6 and GLM 5.2 as selectable models inside its Claygent research agent on July 16, 2026, offered as a lower cost option. Checking which model each of your tools is set to use is the practical first step, since defaults are rarely reviewed after setup.
There is no restriction in force, so this is a preparation question rather than a compliance one. The reasonable position is to know which model each workflow uses, distinguish self hosted weights from a hosted API run by a foreign vendor because the exposures differ, and configure a tested fallback model for every automation so any future change is a setting rather than a rebuild.