Anthropic posted "Our position on open-weights models" on July 27, 2026 in response to active policy discussion around open-weights and U.S. officials considering restrictions. In the post, Dario Amodei says Anthropic has not advocated banning open-weights models.
The post distinguishes open-weights models by risk profile and argues that models without dangerous capabilities are a public good because they help businesses and researchers without adding proprietary lock-ins. It also frames blanket protective bans as a poor fit for what it calls the deeper national-security concern: harmful capabilities, not simple openness.
Anthropic's policy section makes that same distinction operational in practice. The company labels its Usage Policy as applying to everyone who can submit inputs to Anthropic services, including reseller and passthrough access paths.
The policy is explicitly layered by risk: universal standards for all users, dedicated high-risk requirements for elevated-harm use cases, and additional guidelines for specific contexts such as consumer-facing chatbots, minors, agentic uses, and other sensitive patterns.
The Usage Policy page also says Anthropic updates policy as technology and risk patterns evolve, which matters for product teams planning long-horizon rollout decisions around these same areas.
The Consumer Terms of Service update (October 8, 2025) supplies the commercial contract layer that frames what users agree to when they use Anthropic products and services, alongside the safer-use commitments in policy pages.
On the product side, Anthropic's model docs highlight explicit model versioning and lifecycle controls: model IDs, choosing models, migration guidance, and model deprecations. That signals a product path where governance language is tied to concrete release operations and not only public statements.
For policy reporting on this beat: Anthropic did not publish an independent benchmark or cost comparison for these governance claims in the sourced pages, so those figures are not estimated here.
