On 27 July 2026 Anthropic published a post by chief executive Dario Amodei setting out the company’s position on open-weights models. The stated purpose was to correct a characterisation of its policy advocacy: “Anthropic has never advocated for a ban on open-weights models.” The post describes open-weights models without dangerous capabilities as a public good, noting that they cost nothing to use beyond the compute needed to run them.
The post separates that from three measures Anthropic does support. First, controls on advanced semiconductor exports to China, which Amodei describes as the most efficient and direct way to address his primary national security concern, along with enforcement against chip smuggling and workarounds. Second, restrictions targeted at industrial-scale distillation operations, which he distinguishes explicitly from a broad ban on open weights, on the argument that distillation is what allows a competing frontier to close to within a few months of the US frontier. Third, mandatory safety testing for all sufficiently capable models, open and closed alike, with risks determined empirically through testing rather than assumed in advance by category. The post states that protectionist bans would not address his most serious national security concerns.
The publication came amid an active policy dispute over whether the US should restrict open-weight models, including models released by Chinese developers, and after other parts of the industry had publicly defended open weights.
For a business leader, the operative detail is the third measure. A regime of mandatory safety testing keyed to capability rather than to licence type would apply to open-weight models an enterprise self-hosts as well as to hosted commercial APIs. Organisations whose AI strategy depends on running open-weight models internally, often precisely to avoid vendor and compliance overhead, should not assume that open weights sit outside the evaluation requirements now being drafted.