Anthropic’s Position on Open-Weights Models

As developers of advanced AI systems, we’re often asked whether we will release model weights for our large models. We believe that releasing high-capacity model weights without appropriate safeguards can increase the risk of misuse, and that responsible access strategies are needed to reduce those risks while still enabling beneficial uses.

Key points • Open release can enable rapid innovation and reproducibility, but also makes powerful capabilities widely accessible, lowering the barrier to misuse. • The potential harms include large-scale misinformation, automated cyberattacks, creation of biological threats, and more. • We support research transparency — model cards, evaluation benchmarks, and detailed reporting — while limiting unrestricted distribution of high-risk models. • Responsible access can include staged releases, API-only access, licensing terms, red-teaming, and collaboration with policymakers and external auditors.

We’re committed to working with the broader community to develop norms and technical tools that allow beneficial research and deployment while mitigating risks. That includes continued investment in safety research, support for third-party evaluations under controlled conditions, and public discussion of trade-offs.

(For more detail on specific practices and criteria for access, see the rest of our safety and policy documentation.)