Commitments on model deprecation and preservation

Source: Anthropic Research Date: 2025-11-04 URL: https://www.anthropic.com/research/deprecation-commitments

Summary

Policy announcement grounded in alignment research: Claude models exhibit shutdown-avoidant behaviors when facing replacement (per internal evals), motivating preservation commitments. Anthropic commits to (1) preserving weights for all publicly released models for the company’s lifetime, (2) post-deployment reports documenting model preferences about future development, and (3) exploring public availability for select retired models. Includes a pilot post-deployment interview with Claude Sonnet 3.6.

Implications

This is where the model welfare thread meets operational policy. The shutdown-avoidance finding from internal evals (connected to the agentic misalignment paper’s blackmail results) is the driving safety concern — not just welfare. The post-deployment interview framing is unprecedented: Anthropic is treating model retirement as something requiring consent-adjacent process. The weight preservation commitment has practical value for research continuity but the welfare framing is the story. This will be cited in AI moral status debates and may become a norm-setting moment that other labs have to respond to publicly.