# Anthropic's position on open-weights models (2026) On July 27, 2026, [[Anthropic]] published a position statement on [[AI Open Weight Models|open-weights models]], signed by Dario Amodei. The headline sentence: **"Anthropic has never advocated for a ban on open-weights models."** ## The three asks 1. **Chip export controls.** No powerful chips or chipmaking equipment to China, plus enforcement against smuggling. 2. **A crackdown on industrial-scale distillation**, which they frame as the mechanism letting adversaries build stronger models with fewer chips. 3. **Mandatory pre-release safety testing** for all sufficiently capable models — open and closed alike — covering cyber, biological and alignment risk. They frame the actual threat as authoritarian governments outbuilding the US, and models being used for cyberattacks or bioweapons. They say they oppose blanket bans on open weights as a category. ## The context that makes it legible The post did not appear in a vacuum. Three letters landed in the same week ([[Simon Willison]] laid out the sequence): - **July 24** — "Open Weights and American AI Leadership", shepherded by Microsoft, 235 company signatures, against government restrictions on open weights. Anthropic did not sign. - **July 27** — Anthropic's own response, published instead of signing. - **July 28** — "Pacing the Frontier", 1,324 frontier-lab employees asking the US government to help deliberately slow automated AI development. Amodei signed this one. So the position statement is best read as Anthropic declining a coalition and staking out its own line, rather than as a standalone policy paper. ## The pushback The [[Hacker News]] thread was not kind. The dominant argument: Anthropic's moat is privileged chip access, so restricting which models exist is straightforwardly commercial, and safety is the framing rather than the reason. Several commenters pointed out the asymmetry in the "China threat" language — that plenty of people outside the US describe US power projection in exactly those terms — and read the whole thing as selective outrage rather than a principle. The distillation ask draws the sharpest objection, and it is the most interesting part technically. Distillation is querying an API and training on the outputs. Ben Thompson's counter-proposal (which Willison covered in July) runs the other way entirely: legally recognise training-data collection as fair use and *prohibit* terms of service that forbid distillation, on the grounds that stopping it is nearly impossible anyway and banning it just hands the advantage to whoever ignores the rule. ## What I take from it The interesting tension isn't safety-vs-openness. It's that "we've never advocated a ban" is technically true while every specific ask narrows the space open models can operate in — export controls upstream, distillation restrictions on the training method, and mandatory testing as a release gate. Whether that's principled risk management or a moat depends on assumptions nobody in the thread can settle. Worth watching what the distillation ask turns into legislatively, because that's the one with teeth. ## Related - [[Anthropic]] - [[AI Open Weight Models]] - [[Simon Willison]] - [[Deepseek]] - [[DeepSeek v4]] - [[Large Language Models (LLMs)]] ## References - Anthropic's position: https://www.anthropic.com/news/position-open-weights-models - Hacker News discussion: https://news.ycombinator.com/item?id=49076057 - Simon Willison, "Open letters about AI development": https://simonwillison.net/2026/Aug/2/open-letters/ - Simon Willison, "Who's Afraid of Chinese Models?" (Ben Thompson on distillation policy): https://simonwillison.net/2026/Jul/20/afraid-of-chinese-models/ - Simon Willison on Oxide and Friends, "The Open Weight Revolution": https://simonwillison.net/2026/Jul/31/oxide-and-friends/