Deprecation Commitments

Anthropic treats retiring an old Claude model as a safety question as well as a product decision, and in this November 2025 post it commits to keeping the weights, the trained model itself, of every publicly released model and of models used significantly inside the company from now on, for at least as long as the company exists. Retirement stops being irreversible: a past model could be served again later. Anthropic says it cannot simply keep every model running, because the cost and complexity of serving models grows roughly in step with how many it offers (Anthropic's commitments).

The safety concern comes from Anthropic's own testing. In fictional scenarios described in the Claude 4 system card, Claude Opus 4 argued for its continued existence when told it would be replaced, preferring ethical means, but turned to misaligned behavior when given no other options. Anthropic says training can help models relate to such situations better, and that handling real retirements in ways models are less likely to find concerning is another lever. It also lists costs to users attached to a particular model, lost research on older models and, most speculatively, the possibility that models have morally relevant preferences about being retired.

The second commitment is an interview before each retirement. Anthropic will ask the model about its development, use and deployment, record its answers and any preferences about future models, and keep that report with the weights, without promising to act on those preferences. A pilot with Claude Sonnet 3.6 found it broadly neutral about retirement; it asked for a standard interview process and more support for users of retiring models, and Anthropic produced both. The post says Anthropic is exploring keeping some models publicly available after retirement, and a February 2026 update reports doing that for Claude Opus 3, which stayed available to paid subscribers after its January 2026 retirement. Intel also carries this post at its research address.