Responsible Scaling Policy

The Responsible Scaling Policy is a commitment structure: as models get more capable, specified safeguards become mandatory before deployment. Capability thresholds are defined in advance; safeguards are attached to them.

Where it becomes visible

The RSP is not an abstraction if you are watching how models ship:

• System cards report the evaluations that determine which level applies.
• Claude Mythos 5 is invitation-only and scoped to defensive cybersecurity — an availability decision that reflects capability gating rather than commercial strategy.
• Claude Sonnet 5 is described as having substantially reduced cybersecurity capabilities relative to Opus models — a deliberate property.

When a model is powerful but narrowly released, the RSP is usually the reason.

What it is not

• Not the Usage Policy. The RSP constrains what Anthropic deploys. The usage policy constrains what you do with it.
• Not a certification you inherit. Anthropic meeting its own commitments says nothing about whether your deployment is safe.

What this page does not claim

The current safety-level definitions, thresholds and required safeguards are not reproduced here. The policy is versioned and has been updated more than once; read the current text at the source.