Responsible scaling levels (ASL)

Anthropic's Responsible Scaling Policy defines AI Safety Levels (ASLs): tiers of safety and security measures a model must be deployed and protected under, based on how capable it is in high-risk areas.

Why it matters. A model's ASL tells you which safeguards Anthropic committed to for it, which can show up as classifiers that block some requests.

How it works

Anthropic's Responsible Scaling Policy (RSP) is its framework for managing catastrophic risks as models become more capable. It defines AI Safety Levels, loosely modeled on the biosafety levels used for handling dangerous pathogens, where higher levels require stricter deployment and security safeguards.

Anthropic evaluates models against capability thresholds in areas such as chemical, biological, radiological and nuclear (CBRN) weapons and autonomous AI research and development. When a model may cross a threshold, it must be deployed under the matching ASL standard; ASL-3 safeguards, for example, include targeted classifiers and stronger protection of model weights. System cards report the ASL determination for each model.

For builders, a higher ASL can mean classifiers that block some requests in sensitive domains. It describes required safeguards, not quality.

Worked example

Illustrative: a system card states a model is deployed under the ASL-3 Standard. A company building biology research tools takes that as a cue to test its real prompts against the model and check what the API returns when a classifier blocks a request.

Common mistakes

Reading a higher ASL as 'better' or 'more dangerous to use' rather than as a level of required safeguards.
Not testing domain prompts (biology, chemistry, security) against the classifiers that come with higher levels.
Assuming another maker's levels or thresholds map one-to-one onto ASLs.

See it on a real card: Safety & risk on Claude Opus 5.5.