A responsible scaling policy sets out, in advance, what a developer will do as its models become more capable: which dangerous capabilities it will test for, what safeguards it will require at each level, and what it will do if the safeguards are not ready. Anthropic introduced the idea in 2023, and other labs published similar frontier safety frameworks.
The value of such a policy depends on how binding its commitments are. Version 3.0 of Anthropic’s policy, which took effect on 24 February 2026, dropped the earlier implied commitment to pause development when risks were too high. The company’s reasoning was that catastrophic risk depends on the actions of all frontier developers, and that slowing alone could let less careful actors lead.
For a buyer of AI, the policy is evidence worth reading, and worth re-reading when it changes.