Anthropic CEO Dario Amodei published a rebuttal on the company’s blog stating that Anthropic has never advocated banning open-weights models, directly answering an accusation aimed at the company this week. The response follows a letter organized around Nvidia and signed by 77 technology companies urging Washington to leave open-weights models unrestricted, a letter that accused labs like Anthropic of quietly wanting the opposite. Amodei’s post names the accusation and denies it in one sentence: Anthropic has never called for such a ban.
Amodei goes further, calling open-weights models that lack dangerous capabilities a public good, since they cost users only the compute to run them and give developers and researchers a free option. That framing matters because Anthropic sells closed models for a living, and a company staking out generosity toward free alternatives to its own product is a claim worth testing against what it actually proposes to regulate.
The proposal splits into three measures. First, block the sale of advanced chips and chipmaking equipment to China and prosecute the smuggling networks that route around existing controls. Second, target industrial-scale distillation operations, the compute-efficient shortcut that lets a country with fewer chips train competitive models by copying outputs from frontier systems. Third, require mandatory safety testing for cyber, biological, and alignment risk on any sufficiently capable model, whether its weights are open or closed.
Only the third measure applies to Anthropic itself. Chip export controls constrain a rival’s supply chain, not Anthropic’s, since Anthropic is not the one buying restricted hardware from Chinese vendors. A distillation crackdown protects the labs whose outputs get copied, and Anthropic’s own frontier models sit squarely in that category, so the rule shields Anthropic as much as it burdens the operations doing the copying. Mandatory testing is the one proposal that would put Anthropic’s own closed releases through the same pre-release screening it wants applied to open-weights competitors. That is the strongest evidence the position is not pure self-interest dressed up as safety policy, even if the other two measures cost Anthropic nothing.
Amodei frames his two core worries as separate from the open-versus-closed debate entirely. The first is that an authoritarian government, principally the Chinese Communist Party, trains a model more powerful than anything the US has and uses it for military advantage or domestic repression. The second is that a powerful model, regardless of who releases it, gets misused for cyberattacks or bioweapons design. He cites Vice President Vance’s warning at last year’s Paris AI summit and the Intelligence Community’s own 2026 Annual Threat Assessment as evidence the first concern already has traction inside the US government.
On the Nvidia-organized letter itself, Amodei agrees that open weights broaden access and strengthen competition, but rejects its central safety argument, the claim that wide distribution of model weights favors defenders over attackers. He argues the reverse could hold for biological risk: a sufficiently capable model might let someone weaponize a pathogen in weeks, while building a countermeasure takes years, the same mismatch the US saw during Operation Warp Speed. He wants that question settled by pre-release testing rather than argued from priors on either side of the dispute.
The disagreement now has two documented positions instead of one side implying what the other believes. Anyone drafting actual legislation should read Amodei’s three measures against the letter’s asks line by line: the overlap sits at chip controls and distillation enforcement, and the real fight is whether mandatory testing applies evenly to open and closed labs, including Anthropic’s own.
This account is based on Dario Amodei’s post “Our position on open-weights models,” published by Anthropic on its company blog in late July 2026.
