Anthropic has never advocated for banning open-weights AI models, despite recent accusations, CEO Dario Amodei clarified on July 27. Instead, the company supports three specific policy measures: restricting chip exports to China, cracking down on industrial-scale model distillation, and requiring mandatory safety testing for all sufficiently capable models, whether open or closed.

What's Driving the Open-Weights Debate?

The clarification comes amid heated discussion about open-weights models, particularly those developed in China. Some U.S. officials have reportedly considered banning American companies from using Chinese open-weights models. In response, multiple tech companies signed a letter supporting open-weights development, while critics accused Anthropic of wanting to restrict these models to protect its own business interests.

Amodei addressed the misconception directly, stating that Anthropic has consistently opposed blanket bans. "Open-weights models that don't have dangerous capabilities are a public good," he explained, noting that they cost nothing beyond the compute needed to run them and provide value to businesses, developers, and researchers.

Amodei

What Are Amodei's Real National Security Concerns?

Rather than focusing on open-weights as the threat, Amodei outlined two specific nightmare scenarios that concern him. First, he worries that authoritarian governments, particularly the Chinese Communist Party, could build AI models more powerful than those developed by the United States and use them to achieve permanent military superiority or enable deep repression of their own populations. Second, he expressed concern that powerful AI models could be misused for cyberattacks or biological attacks, and may have serious alignment problems.

Amodei noted that these concerns are widely shared within the U.S. government. Vice President Vance warned in Paris that "authoritarian regimes have stolen and used AI to strengthen their military, intelligence, and surveillance capabilities," and the Intelligence Community's 2026 Annual Threat Assessment found that "other global powers' robust progress in AI is challenging US economic competitiveness and national security advantages".

Amodei

How Should Policymakers Address These Risks?

Rather than banning open-weights models, Amodei proposed three targeted measures that he and Anthropic have consistently advocated for:

• Restrict Advanced Chip Access: The U.S. should not sell powerful chips or chipmaking equipment to China and should crack down on smuggling and workarounds used to obtain such chips. Since China has limited domestic production capacity, restricting chip access is the most direct way to prevent authoritarian governments from building more powerful models than the U.S. .
• Crack Down on Industrial Distillation: Distillation is a much more compute-efficient process than training models from scratch, allowing China to build better models than its chip supply would ordinarily enable. While distillation does not allow the Chinese Communist Party to obtain equivalent capabilities to the U.S., it can bring the Chinese frontier to within a few months of the American frontier. Policy interventions should deter this behavior.
• Mandate Safety Testing: All sufficiently capable models, whether open or closed, should undergo mandatory safety testing for cyber, biological, and alignment risks before release. Amodei noted that this approach has gained traction, with both the Trump administration and recent industry proposals moving in this direction.

Amodei emphasized that effective testing would need to be global, meaning even the Chinese Communist Party would need to participate. He suggested this may be possible because preventing AI biological weapons is in China's interest as well.

Where Does Anthropic Disagree With the Open Letter?

While Amodei agreed with much of the open letter supporting open-weights models, he disagreed with some of its core assertions. The letter claimed that open-weights models necessarily make it easier to develop safeguards and that broad access to capabilities helps defenders more than attackers. Amodei argued the opposite may be true.

"I worry that biology will have a strong attacker-defender asymmetry, where 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," Amodei stated.

Dario Amodei, CEO at Anthropic

He argued that questions about whether open-weights models pose increased risks should be empirically answered through rigorous pre-release testing, not assumed in advance.

What Does This Mean for Claude's Development?

While Amodei's policy statement focused on the open-weights debate, Anthropic has been making significant technical advances in its Claude models. The company recently removed more than 80 percent of Claude Code's system prompt for its newest models, Opus 5 and Fable 5, without any measurable loss in performance across internal coding evaluations.

This change, disclosed on July 24, moved behavioral guidance into tools and selectively loaded skills rather than keeping everything in a single instruction block. The reduction suggests that newer Claude models possess better implicit behavioral understanding and need less explicit instruction to behave correctly. Older Claude models, including Sonnet 4.5 and Haiku, still require the full system prompt.

The 80 percent reduction challenges a core assumption in AI safety engineering: that more explicit constraints produce safer behavior. If that assumption held true, removing 80 percent of constraints should have produced measurable increases in misalignment or refusal anomalies. Anthropic reported none of that occurred.

For developers using Claude Code daily, the change means faster responses and lower costs, since system prompt tokens count against both context windows and billing. The reduction also decreases unexpected refusals that sometimes occur when one behavioral instruction accidentally cancels or confuses another.

Amodei's position on open-weights models reflects Anthropic's broader philosophy: focus on the real threats, use targeted policy measures, and trust that better models with stronger implicit understanding require fewer explicit guardrails. Whether that confidence is justified remains an open question for external researchers and policymakers.