Anthropic Draws a Line: Inside Its 'Position on Open-Weights Models'

On Monday, July 27, 2026, Anthropic CEO Dario Amodei published a post titled Our position on open-weights models, and led with an emphatic denial: Anthropic has "never advocated for a ban on open-weights models." The announcement on X pulled in roughly 4.8M views within hours.

The trigger was an open letter, Open Weights and American AI Leadership, signed the previous Friday by around 25 companies including Nvidia, Meta, and Microsoft. Over the weekend OpenAI, Google, and SpaceX added signatures, which left Anthropic as the only major lab that hadn't signed. Critics read that silence as quiet support for a crackdown, and the post is Amodei's attempt to say otherwise.

The post opens with a concession, not a warning. Open-weights models that lack dangerous capabilities are "a public good," Amodei writes: they cost only the compute to run them, and they help businesses, developers, and researchers. The disagreement is entirely about the capable end of the spectrum.

He frames two nightmare scenarios. The primary one is an authoritarian government, with China named as "clearly the most capable threat," building AI more powerful than the US's for military dominance or domestic repression. The secondary one is a capable model used for cyber or biological attacks, where open weights are harder to guardrail and, once out, cannot be recalled.

Instead of a ban, Amodei lists three measures he says Anthropic has "consistently advocated for." The point is to slow a hostile frontier without shutting off the benign open-weight ecosystem.

The third ask is the load-bearing one: mandatory safety testing for all sufficiently capable models, open and closed, with exemptions for smaller models from startups and academia. Amodei calls this "close to a consensus" and points to Trump administration movement in that direction.

The industry letter rests on a familiar claim: broad access to model weights helps defenders more than attackers and makes safeguards easier to build. Amodei disputes both halves of that directly.

His example is bioweapons. A capable model could help someone weaponize a pandemic-scale virus fast, while building a defense is, using the Operation Warp Speed precedent, "a multi-year operational task in the best case." The asymmetry, he argues, is the whole problem: reporters described the post as Amodei doubling down on risk while still denying he wants a ban.

The critics were loud. David Sacks, Bill Gurley, and Kai-Fu Lee accused Anthropic of wanting restrictions to protect its commercial position, with Lee noting that who didn't sign the letter was more interesting than who did.

Matthew Berman's rebuttal captured the sharper objection: Anthropic hasn't asked for a ban, but as one of the largest labs, its steady messaging about open models being unsafe functions as advocacy for one anyway. The words carry weight even when they stop short of a demand.

There is also the awkward backdrop: Anthropic releases no open-weight models of its own. One analysis reports that in June 2026 the US ordered Anthropic to bar foreign nationals from its Fable 5 and Mythos 5 models, and the company shut both off worldwide rather than police nationality, described as the first export measure aimed at models rather than chips (single-source, treat as reported). Separately, Platformer reported that cybersecurity experts were rattled by a Mythos model's offensive capabilities.