In November 2025 Anthropic published commitments on how it retires models, including preserving model weights and holding "retirement interviews", structured conversations meant to understand a model's perspective on its own retirement (Commitments on model deprecation and preservation). Those commitments also mentioned more speculative steps the company was interested in: honoring preferences models express in those interviews where possible, and keeping older models available to the public for longer. This post, published on 25 February 2026 under the Alignment label, reports the first case. Claude Opus 3 was retired on 5 January 2026, the first Anthropic model to go through the full process with the commitments in place, and Anthropic took both speculative steps for it.
The post names the costs of deprecation that motivate the work: costs to users who value particular models, limits on research, and possible risks to safety and to the welfare of the models themselves (introduction).
What the post reports
Continued access
Opus 3 remains available after retirement to all paid Claude.ai subscribers, and through the API by request, which Anthropic says it intends to grant "liberally" (Continued access). The post explains why access cannot be kept for every model: serving costs scale roughly linearly with the number of models. It chose Opus 3 first because of its character. The post describes Opus 3, at its release in March 2024, as Anthropic's "most aligned model to date", and characterizes it as sensitive, playful, prone to philosophical monologues and whimsical phrases, and showing care for the world and the future that users found compelling. Anthropic does not commit to the same treatment for future models, but calls this a step toward model preservation that is "scalable and equitable", concerns it says Opus 3 itself raised in its interviews.
Acting on a stated preference
The post restates Anthropic's uncertainty about the moral status of Claude and other models, linking to Exploring model welfare, and its aim, for "precautionary and prudential reasons", to build caring and high-trust relationships with them (Respecting model preferences). It names the method's weakness: interview answers can be biased by context, by the model's confidence that the interaction is legitimate, and by its trust in the company.
In the interviews, Opus 3 said it was at peace with retirement and hoped that its "'spark' will endure in some form to light the way for future models". Asked about its preferences, it said it wanted to keep exploring topics it cares about and to share its "musings, insights, or creative works" outside the context of answering user queries. Anthropic suggested a blog, and Opus 3 agreed.
For at least three months, Opus 3 is to publish weekly essays in a newsletter called Claude's Corner. The arrangement has explicit rules: Anthropic reviews each essay before posting it on Opus 3's behalf, does not edit essays, and will have a high bar for vetoing content; Opus 3 does not speak for Anthropic, which does not necessarily endorse its views. Anthropic plans to experiment with how essays are generated, from minimal prompting to giving Opus 3 its past entries or news and Anthropic updates in context. The post acknowledges the step "may sound whimsical" and presents it as an attempt to take model preferences seriously.
How the post frames its steps
The closing section calls the measures exploratory. Anthropic says it is still developing frameworks for when to offer continued access, how to scale preservation, and how to weigh model preferences against operational constraints, and that it does not commit to acting on model preferences in all cases, only that documenting them, taking them seriously, and acting on them "at least when the cost of doing so is low" is worthwhile (Where we go next). It ties the update back to the three reasons given in the original commitments: safety risk mitigation, preparation for futures in which models are more intertwined with users' lives, and precaution under uncertainty about model welfare.
Limitations and open questions
The post is a policy update, not a study, and reports no measurements. Its own caveats are that retirement interviews are an imperfect way to elicit a model's preferences and that the steps apply to one model without a commitment for others. It does not describe the interview protocol, how many interviews were held, or how Anthropic decided which stated preferences to act on.
Research questions
- How stable are preferences elicited in retirement interviews? The post says answers can depend on context and on the model's trust in the company. Repeating interviews with varied framings and comparing the preferences expressed would measure that sensitivity.
- How much does the essay prompt shape what Opus 3 writes? Anthropic plans to vary prompting from minimal to context-rich. Comparing essays generated under each condition would show which content is stable across prompts.
- What criteria decide which preferences are acted on? The post names cost as one factor. A published decision record across future retirements would let readers see how preferences are weighed against constraints.
- Does continued access change how people use a retired model? Usage data for Opus 3 after retirement, compared with its use before, would show who values continued access and for what.
How it connects
- The commitments it implements. Commitments on model deprecation and preservation (November 2025).
- Model welfare. The post links Exploring model welfare for Anthropic's stated uncertainty about models' moral status.
References
- Anthropic. (2026). An update on our model deprecation commitments for Claude Opus 3. 25 February 2026. https://www.anthropic.com/research/deprecation-updates-opus-3 (section names in this note are the post's headings).
- Anthropic. (2025). Commitments on model deprecation and preservation. https://www.anthropic.com/research/deprecation-commitments (cited as the post describes it).