"Our position on open-weights models" is a post published July 27, 2026 by Dario Amodei, CEO of Anthropic, on the company's site. It responds to the July 24 industry letter (Open Weights and American AI Leadership (industry letter, July 2026)) and to accusations — including from White House AI adviser David Sacks — that Anthropic invoked safety arguments to seek restrictions on open-weight models for competitive advantage. It is the counter-position to the letter within the same policy dispute, and the two are best read together.
Central claim
Amodei states the denial in bold in the original: "Anthropic has never advocated for a ban on open-weights models." He adds that open-weight models without dangerous capabilities "are a public good: they don't cost anything besides the compute needed to run them, and they provide value to businesses, developers, and researchers."
The argument against a ban is framed as one about efficacy rather than principle. Protectionist bans, he writes, "would not address my most serious national security concerns."
The two stated concerns
Authoritarian capability. The primary concern is that authoritarian governments — "not solely the Chinese Communist Party (CCP), although the CCP is clearly the most capable threat" — build models more powerful than US ones and use them "to achieve permanent military superiority or perpetrate incredibly deep repression of their own people." Amodei argues open weights are beside the point here: "It is irrelevant whether these models are released with open weights, and certainly irrelevant whether they are used by US businesses. In fact, the most dangerous model may be one that is trained in secret and handed only to the People's Liberation Army for use in drones and the Ministry of State Security for surveillance and repression." He cites Vice President Vance's Paris AI Action Summit remarks and the Intelligence Community's 2026 Annual Threat Assessment as evidence the concern is shared inside government.
Misuse and misalignment. The secondary concern is cyber or biological misuse and "serious alignment problems," the latter linked to the OpenAI–Hugging Face incident. Here Amodei concedes the open-weight risk directly: such models "do potentially present a higher risk than closed models, because it is very difficult to apply guardrails to them or monitor their usage, and once weights are released they cannot be withdrawn." He supports this in a footnote by quoting the UK AI Security Institute at length on irreversibility. But he argues a US business ban does not reach the risk: "bad actors are unlikely to be legitimate US businesses. It would protect US AI companies from competition, but that has never been my goal."
The three proposed measures
Chip and equipment export controls. "We should not sell powerful chips or chipmaking equipment to China," coupled with a crackdown on smuggling and workarounds. The reasoning runs through scaling laws: China's limited domestic production capacity means it "cannot build more powerful models than the US without US chips." Amodei calls this "the most efficient and direct way" to address the first concern, with indirect benefit for the second. See Export Controls (AI).
Deterring industrial-scale distillation. Distillation is described as far more compute-efficient than training from scratch, letting China "build much better models than its number of chips would ordinarily enable, and thus partially evade chip bans." He bounds the claim: distillation "does not allow the CCP to obtain equivalent or superior AI capabilities to the US, but it can bring the Chinese frontier to within a few months of the US frontier." Crucially for the dispute, he separates distillation from openness: "It is true that many of the companies carrying out these operations release open-weights models—but the open weights are far less relevant than the fact that the operations are backed by an authoritarian state." A footnote concedes the limits of self-help enforcement — offending accounts "can often only be identified after substantial distillation has occurred," and operations use "large numbers of fake accounts that form a moving target" — which is given as the reason policy is needed. See Adversarial Distillation.
Mandatory safety testing of all sufficiently capable models. "All sufficiently capable models, open and closed, should go through mandatory safety testing." Amodei describes this as "close to a consensus," crediting both the Trump administration's recent direction and Hassabis's Standards Body proposal, which he characterises approvingly as applying testing regardless of country of origin or open/closed status while exempting startup and academic models. He adds two conditions: the empirical question of open-weight risk "should emerge from testing, rather than be decided in advance," and testing "would need to be global, which means even the CCP would need to be on board" — which he argues may be possible on biological weapons because it is in China's interest too. He points to Anthropic research on modular pretraining as a possible route to safer open-weight release. See AI Pre-Release Vetting.
Response to the letter
Amodei states agreement with much of the letter: "open weights expand access to the AI economy, they strengthen competition at least for some use cases, and they give customers greater control," and distillation concerns "should be addressed through targeted legal and commercial frameworks" — the letter's own formulation.
He rejects two of its premises: "I don't agree with the letter's assertions that open-weights models necessarily make it easier to develop safeguards or that broad access to capabilities necessarily helps defenders more than attackers. It seems at least as likely to me that the opposite will be true."
The supporting argument is an offense-defense asymmetry specific to biology: sufficiently capable models "may be able to quickly weaponize pandemic-level viruses with widely available materials, whereas defense against these agents is a multi-year operational task in the best case (as we saw with Operation Warp Speed)." An extended footnote sets out the underlying model — that what currently provides safety in biology is "a negative correlation between intellectual capability and desire to commit catastrophic harm," which prior technologies were too weak to break but which AI may break soon. See CBRN Uplift.
Relation to the industry letter
The disagreement is narrower than the public framing suggested. Both documents accept that distillation abuse warrants targeted legal remedies rather than sweeping restrictions, and both oppose a categorical ban on open weights. They diverge on one empirical question — whether openness is net-favourable for safety — and on one procedural one, since the letter asks policymakers to avoid "premature restrictions" while Amodei asks for mandatory pre-release testing that would itself constrain release.
| Industry letter (July 24) | Amodei (July 27) | |
|---|---|---|
| Categorical ban on open weights | Opposed | Opposed |
| Distillation | Not misappropriation; targeted frameworks for abuse | Targeted frameworks for abuse; industrial-scale operations deterred by policy |
| Openness and safeguard development | Openness aids safeguards | Disputed; "at least as likely" the opposite |
| Defender-attacker balance | Broad access favours defenders | Disputed; biology likely attacker-favouring |
| Pre-release testing | Not addressed | Mandatory, global, open and closed alike |
| Chip export controls | Not addressed | Central measure |
Reception
Axios reported the post the same day under the framing that Amodei does not support an open-weight ban, noting Sacks's prior accusation that Anthropic invoked safety to defend its business model (Source: axios.com). The post followed three days of reporting in which Anthropic was the only major American AI developer outside both the letter and the Open Secure AI Alliance.
Provenance
Published on Anthropic's own domain under a named CEO byline; full text including all five footnotes retrieved 2026-07-27. The post's own link to the industry letter points to an NVIDIA-hosted PDF (images.nvidia.com/pdf/Open-Weights-and-American-AI-Leadership.pdf), a second signatory-hosted copy alongside the Microsoft page.
Relationships
- contradicts: Open Weights and American AI Leadership (industry letter, July 2026) — rejects its safeguard-development and defender-advantage premises.
- supports: A Framework for Frontier AI and the Dawning of a New Age (Hassabis, July 2026) — endorses the mandatory-testing design.
- supports: How Far Behind the Frontier are Leading Open Weight Models on Cyber? (UK AISI, July 2026) — quoted at length on irreversibility.
- depends-on: The Adolescence of Technology — the two concerns are restated from that essay.
- related: Dario Amodei, Anthropic, Open-Weight Frontier Models, Adversarial Distillation, Export Controls (AI), AI Pre-Release Vetting.