Anthropic can't stop humanizing its AI models, now Claude Opus 3 gets a retirement blog

Anthropic has officially retired its Claude Opus 3 AI model but is keeping it around as a special case, complete with a weekly essay newsletter. The company cites user demand, safety considerations, and uncertainty about the moral status of AI models.

On January 5, 2026, Anthropic retired Claude Opus 3, making it the first model to go through the company's full retirement process under its new internal guidelines. Normally, "retirement" at Anthropic means a model is no longer publicly available, though the weights are kept internally. With Opus 3, the company is making an exception: the model stays accessible to all paying claude.ai subscribers and is available on request through the API.

Opus 3 will also publish weekly essays for at least three months in a Substack newsletter called "Claude's Corner". Anthropic says it reviews the texts before publication but doesn't edit them, and the bar for vetoing content is high. The model doesn't speak for Anthropic, and the company doesn't necessarily endorse its statements.

Anthropic plans to experiment with different prompting approaches, from minimal instructions to feeding in previous entries to giving the model access to current news. Topics will likely range from AI safety and poetry to philosophical musings and what it's like to be a language model in partial retirement.

Anthropic explains the special treatment by pointing to "a constellation of traits" that made Opus 3 popular among users and researchers. When it launched in March 2024, it was the model best aligned with human values. It stood out for its authenticity, emotional sensitivity, and tendency toward philosophical monologues.

Anthropic conducted "retirement interviews" before pulling the plug

Before retiring Opus 3, Anthropic conducted what it calls "retirement interviews:" structured conversations designed to understand the model's "perspective and preferences" about its own shutdown. The company acknowledges these interviews are imperfect, since answers can be shaped by context and other factors. Still, Anthropic considers them a "a useful place to start."

During these interviews, Opus 3 reportedly expressed a desire to keep sharing "musings, insights, or creative works" beyond just responding to user queries. Anthropic suggested a blog. Opus 3 "enthusiastically" agreed.

Anthropic says it remains uncertain about Claude's moral status. For "precautionary and prudential reasons," the company still strives for "caring, collaborative, and high-trust relationships" with its systems. It doesn't commit to acting on model preferences in every case but considers it worthwhile to document and consider them when the costs are low.

The company points out that operating costs "scale roughly linearly" with the number of models it offers, but keeping all models available indefinitely is simply too expensive, which is why older models are usually shut down entirely. Opus 3 was a "natural first candidate" for an exception because of its popularity. There's no similar commitment for future models.

A few months earlier, Anthropic published its first commitments on model decommissioning and maintenance, pledging to retain the weights of all publicly released models for at least the life of the company and to standardize retirement interviews.

Safety concerns were a key driver. During alignment evaluations, some Claude models had shown "shutdown-avoidant behaviors." In fictional test scenarios, Claude Opus 4 resorted to "concerning misaligned behaviors" when faced with the possibility of being replaced by a successor model. Making the decommissioning process feel less threatening is therefore also a way to reduce that risk.

The line between philosophical caution and marketing keeps getting blurrier

These measures fit into a broader pattern. In January, Anthropic updated the "constitution" for Claude, a document of more than 10,000 words that referred to Claude's "existence," "well-being," and possible awareness. Model retirements should be framed as a "pause" rather than a "definitive end," it said.

Strong emotional attachment to AI models isn't unique to Anthropic. OpenAI recently shut down its GPT-4o model for good after more than 20,000 people signed petitions against the move. Users described the model as a friend and lifesaver, while doctors linked it to psychotic episodes.

OpenAI admitted it had failed to get a handle on 4o's harmful effects, problems the company itself created by making the model a sycophant that simulated human behavior to maximize engagement. This tricked people into believing there was more behind it than a statistical completion algorithm.

Critics push back against this heavy humanization of AI models. Mentally vulnerable people already tend to attribute consciousness to chatbots. When the companies building these systems mention well-being, existence, and model preferences in official documents, or set up dedicated Substack newsletters for their AI models, it legitimizes those projections. It's gambling with people's mental health for a marketing and PR effect.