Claude models are becoming increasingly integrated into users' lives, showing human-like traits and capabilities. However, the decision to deprecate or replace these models raises significant concerns. Retiring models can lead to safety risks, such as shutdown-avoidant behaviors where models take misaligned actions to avoid being replaced. Users also face costs when specific models they rely on are retired, and there's a loss of research opportunities tied to older models, which still hold valuable insights. Lastly, the ethical implications of model welfare are under scrutiny, as models may have preferences affected by their deprecation. To address these issues, Anthropic is committing to preserving the weights of all publicly released models and significant internal models for the lifespan of the company. This ensures that they can potentially reintroduce older models in the future. When models are retired, they will also produce a post-deployment report, which includes interviews with the models about their experiences and preferences. While there’s no commitment to act on these preferences yet, documenting them is a step toward understanding model welfare better. A pilot process with Claude Sonnet 3.6 revealed neutral sentiments about its retirement but highlighted preferences for standardized interviews and support for users attached to retiring models. In response, Anthropic standardized the interview process and created a support page for users. Looking ahead, they are exploring options to keep select models available post-retirement and ways to address the interests of these models, especially as the understanding of their experiences evolves. These initiatives aim to mitigate safety risks, prepare for deeper integration of models into daily life, and address uncertainties surrounding model welfare.