Anthropic CEO Dario Amodei clarified on Monday that his company has never advocated for a ban on open-weights models, pushing back against industry criticism following an open letter signed by Nvidia, Microsoft, Meta, and others urging policymakers not to impose broad premature restrictions
on open-weight AI models.
The Open-Weights Schism and Jensen Huang’s X Debut
The debate over downloadable artificial intelligence reached a new inflection point when Nvidia founder and CEO Jensen Huang published his first post on the social platform X to share an open letter. The letter urged policymakers not to impose broad premature restrictions
on open-weight AI models.
Nvidia and a long list of other AI companies including Hugging Face, Meta, Microsoft, and Mistral signed the coalition. Yet the absence of Anthropic drew industry scrutiny, as Anthropic operates as a provider of closed models through its Claude family of products.
Dario Amodei Confronts the Ban Allegations
Critics across the sector raised concerns that Anthropic supports efforts by the U.S. government to ban open-weight Chinese models, or open-weight models generally. Responding to the mounting industry pressure, Anthropic CEO Dario Amodei published a blog post addressing the rumors.
“Anyone who has read my past writing should know that I don’t regard such bans as a useful measure, but let me state it clearly so that there is no doubt: Anthropic has never advocated for a ban on open-weights models.”
Dario Amodei, CEO of Anthropic
While Amodei stated that open-weights models that don’t have dangerous capabilities are a public good, he maintained distinct reservations. He argued that open-weight models are more dangerous in scenarios involving biological attacks or cybersecurity because it’s difficult to apply guardrails to them or monitor their usage.
Moonshot AI and the Threat of Kimi K3
The policy debate unfolded concurrently with a major technical disruption from overseas. Chinese outfit Moonshot AI released the weights for its latest model, Kimi K3, making them freely available for any developer running a contemporary rack of AI hardware. The release demonstrated that open-weight models could closely trail top-tier proprietary systems like OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable while running at a fraction of the cost.
Moonshot’s internal benchmarks indicate that Kimi K3 charges $3 per million tokens for standard non-cached input, compared to $10 for Claude Fable and $5 for Sol. Furthermore, optimized caching structures can drive input costs down to $0.30 per million tokens for repetitive coding workloads. The model achieves this efficiency through a sparse mixture-of-experts architecture and low-precision data types like MXFP4 and MXFP8, allowing it to run efficiently even on lower-end hardware.
Distillation Claims and National Security Concerns
Beneath the commercial competition lies a tense geopolitical struggle over intellectual property and military capabilities. The debate in the industry has centered on allegations that Chinese AI labs are growing in capability often by stealing intellectual property from their American counterparts, including through distillation.
Amodei emphasized that protectionist bans miss the primary national security threat. Instead, he argued that policymakers should focus on restricting adversary access to powerful chips, calling for a formal crackdown on distillation, and instituting global safety testing frameworks that require participation from all capable AI laboratories regardless of origin.
The Economic Argument for Open Architectures
Defenders of open-weight systems view transparency as a prerequisite for long-term economic sustainability. By letting startups, hospitals, factories, and universities download, inspect, and run advanced models locally, organizations avoid depending entirely on a single provider’s pricing roadmap and behavior.
As commercial pressure mounts following the entry of low-cost models from labs like Moonshot, the artificial intelligence sector remains sharply divided over whether the benefits of open architecture outweigh the risks of untraceable modifications and rapid capability diffusion.
