System cards/anthropic

Anthropic

AI safety lab. Claude family of models. Responsible Scaling Policy commits to specific evaluation thresholds before deploying or training models of given capability levels.

Canonical site ↗Seoul signatoryParis signatory
Framework versions (Responsible Scaling Policy)
  1. 2025-03-31RSP v2.3pending

    Incremental v2.3 update. Pending verification of exact change set against the published changelog.

    • Pending verification of v2.3-specific changes from the public RSP changelog
    Primary source ↗PDF ↗
  2. 2024-10-15RSP v2.0pending

    Major v2 restructure of the RSP. Reframes capability commitments around specific evaluation thresholds and introduces explicit ASL-3 and ASL-4 deployment-and-training commitments.

    • Restructured around capability thresholds rather than model-version commitments
    • Introduced explicit ASL-3 and ASL-4 commitments
    • Added bioweapons and autonomy-replication threshold criteria
    • Required Responsible Scaling Officer role
    Primary source ↗PDF ↗
Model cards
  • 2025-09-25against RSP v2.3pending
    Claude Opus 4.5
    Capability: ASL-3
    Claude.ai consumer + API + Enterprise
    CBRNmediumCybermediumAutonomymedium
Other labs tracked