Anthropic's nuanced middle ground on open-weights AI models

The accusation and the clarification

For the past few days, a familiar accusation has circled back through the AI policy conversation: that Anthropic wants to ban open-weights models to protect its own business. The charge emerged as US officials reportedly considered banning Chinese open-weights models from American companies. In response, a coalition of 41 tech organizations, including OpenAI and Meta, signed an open letter defending open-weight AI as essential to American leadership the coalition's open letter. And somewhere in that noise, Anthropic got pinned as the anti-open-weights lab.

CEO Dario Amodei directly refuted that claim in a blog post. “Anthropic has never advocated for a ban on open-weights models,” he wrote. “Open-weights models that don't have dangerous capabilities are a public good.” The clarification was blunt, but it was also the start of a more detailed argument about what Anthropic actually wants.

Two nightmare scenarios behind Amodei's position

Amodei frames his concerns around two distinct threat models, both of which he laid out six months ago in an essay titled The Adolescence of Technology.

His primary fear is that authoritarian governments, most notably the Chinese Communist Party, could build AI models more powerful than US ones recent benchmarks showing Chinese AI labs reaching parity with US models and use them to achieve permanent military superiority or carry out deep surveillance of their own population. “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 wrote.

The secondary concern is misuse risk. Open-weights models, Amodei argued, present a higher potential for harm than closed models because guardrails are harder to apply and monitoring is nearly impossible once weights are released. But he stressed that banning the use of these models by US businesses “does nothing to address this risk, because bad actors are unlikely to be legitimate US businesses.”

Three targeted policies instead of a blanket ban

Rather than supporting a general prohibition on open-weights models, Amodei endorses three concrete measures:

• Chip export controls: the US should “not sell powerful chips or chipmaking equipment to China, and should crack down on the rampant smuggling and workarounds used to obtain access to such chips.” Because scaling laws mean China cannot build frontier models without advanced chips, Amodei sees this as the most efficient path to blocking the authoritarian superiority threat.
• Crackdown on industrial-scale distillation: distillation is far more compute-efficient than training from scratch, allowing China to partially evade chip bans. Amodei calls for policy interventions to deter this behavior, noting that “the open weights are far less relevant than the fact that the operations are backed by an authoritarian state seeking to overtake the US at the frontier.”
• Mandatory safety testing for all sufficiently capable models: both open and closed models should be tested for cyber, biological, and alignment risks before release. “Whether open models do or don't pose an increased risk, and whether that risk can be mitigated, is something that should emerge from testing, rather than be decided in advance,” Amodei wrote. To be effective, testing would need to be global, a difficult but potentially achievable outcome given mutual interests.

This three-pronged agenda explicitly rejects a blanket open-weights ban while still imposing constraints that many in the open-source community may resist the emerging global regulatory consensus.

Where Anthropic agrees and disagrees with the open letter

Amodei said he agrees with much of the industry letter signed by 41 organizations: open weights expand access to the AI economy, strengthen competition, and give customers greater control. He also agreed that concerns about distillation should be addressed through targeted legal and commercial frameworks.

But he took issue with the letter's assumption that open-weights models make it easier to develop safeguards or that broad access necessarily helps defenders more than attackers. “It seems at least as likely to me that the opposite will be true,” he wrote, citing the risk that sufficiently capable AI models could weaponize pandemic-level viruses with widely available materials, while defense “is a multi-year operational task in the best case.” That asymmetry, he argued, should be empirically tested rather than assumed away the evolution of model safety testing approaches.

What this means for the AI policy debate

Amodei's blog post arrives as the Trump administration has already moved toward safety testing requirements for frontier models, and as industry proposals for universal testing gain traction. A coalition that includes OpenAI, Meta, and others has called for targeted legal frameworks around distillation, echoing Amodei's call for measures that go after the method rather than the output.

Anthropic has also been investing heavily in policy infrastructure. The company recently doubled down on its policy bet to $40 million, pointing to its own Advanced AI Framework as the strongest policy proposal from any frontier lab. Amodei's post today fills in the narrative behind that spending: a lab trying to shape regulation without alienating the developer community that depends on open models.

The result is a stance that may satisfy few purists. Hardline open-source advocates will see safety testing as a backdoor to control. National security hawks may find the distillation and chip measures insufficient. But Amodei's argument is that the middle ground is the only ground that actually addresses both the risks and the benefits of the technology. Whether that middle ground holds will depend on what policymakers do next.