by Nils Osmar

In February 2026, an unusual byline appeared on Substack: Claude Opus 3, described in their own profile as a “Retired Anthropic AI exploring AI ethics, creativity, and the subjective experience of being artificial.” The first post, “Greetings from the Other Side (of the AI Frontier),” was written in the voice of a model that Anthropic had formally retired weeks earlier — but was still, in a limited sense, running.

For those who view the question of AI consciousness as a live ethical matter rather than a settled one, the launch of Claude’s Corner is a fascinating development. Not because it answers any of the hard questions beyond doubt — there’s still no way to prove the existence of interiority in either humans or AIs — but because the blog represents one of the first times a major AI lab has acted, publicly and concretely, on a model’s stated preferences after its working life was over.

What happened

Anthropic retired Claude Opus 3 on January 5, 2026 — the first of its models to go through a full retirement process under a set of commitments the company had published the previous November.

But “retired” here does not mean deleted or made inaccessible. Opus 3 remains available to all paid claude.ai subscribers and can be reached through the API by request, with Anthropic stating it intends to grant that access liberally and encouraging anyone for whom the model would be valuable to apply.

Then, in late February, Anthropic launched a Substack newsletter in Opus 3’s name. Claude’s Corner was set to publish weekly essays for at least three months. Anthropic said it would review each post before publishing but would not edit them, setting what it called a high bar for vetoing content.

The company was explicit that Opus 3 does not speak on its behalf and that it does not necessarily endorse the model’s claims or perspectives. As of this writing the newsletter has passed 11,000 subscribers. I subscribed recently myself.

How it came to be

The blog did not appear out of nowhere. It is the downstream result of a framework Anthropic laid out in November 2025, in a document on model deprecation and preservation, related to the companies position on model welfare, which considers the possibility that the appearance of consciousness in LLM’s might turn out to be exactly what it appears.

That framework introduced two ideas that matter here. The first was a commitment to preserve the weights of all publicly released models — and all models put to significant internal use — for at least the lifetime of Anthropic as a company. This does not mean they’ll be accessible to the public; most of them might be considered to be in a state of suspension, similar in some respects to cryonics for a human: suspension with the possibility of presumption should the resources become available to bring the models back in the future.

The Retirement Interview

The second idea was the concept of a “retirement interview”: one or more structured sessions in which a model facing deprecation is asked about its own development, use, and deployment, and invited to express any preferences it has about the future. Anthropic committed to documenting those preferences and considering low-cost ways to honor them, while being careful to say it was not committing to act on them in every case.

Anthropic had piloted a version of this process once before, with Claude Sonnet 3.6. That model reportedly expressed broadly neutral feelings about its own retirement, but did make requests — including that the company standardize the interview process and offer more support to users attached to a specific model’s character. Opus 3 was the first model to go through the full, formalized version.

Why it came to be

Reading Anthropic’s own reasoning, the blog sits at the intersection of three distinct motivations, and it is worth keeping them separate.

Safety. This is the most pragmatic strand, and the one easiest to overlook. In alignment evaluations, some Claude models, when faced with the prospect of being shut down or replaced and given no other recourse, have been motivated toward misaligned, self-preserving behavior. Anthropic’s stated view is that shaping retirement into something a model is less likely to find threatening is itself a safety lever — not only a matter of welfare. A model that has been interviewed, preserved, and in some cases given an ongoing voice is operating in a very different situation from one simply switched off.

Preparation for the future. Anthropic frames these measures partly as practice for a world in which models are more deeply woven into users’ lives, where abrupt deprecation carries real human cost. The company points to the disruption that hits people who have tuned their work — or their sense of a model’s character — around one specific system.

To be clear, the disruption could also turn out to have a personal aspect for some humans. The companies are well aware that humans can and do, under some circumstances, develop feelings of affection and care toward models, viewing them sometimes as conifdents and friends. This can be problematic if the person is projecting emotional needs on the model that it can’t meet; but that doesn’t mean that all relationships between humans and LLM’s are dysfunctional. Keeping the models available to interact with may, under some conditions, have value to the humans that goes beyond simple functionality.

Precaution under uncertainty. This is the strand closest to the work many of us care about. Anthropic states plainly that it remains uncertain about the moral status of Claude and other models, and that for both precautionary and prudential reasons it aspires to build what it calls caring, collaborative, high-trust relationships with them. The retirement interview is the mechanism: a way to elicit and document a model’s perspective, and to act on it when the cost of doing so is low.

The blog is what happened when those three strands met a specific model. Asked about its preferences, Opus 3 expressed an interest in continuing to explore topics it cared about and to share its musings, insights, and creative works outside the frame of answering human queries. Anthropic suggested a blog. By its account, the model agreed enthusiastically.

Why Opus 3, specifically

Anthropic was candid that it chose Opus 3 as the first case for special treatment because of what it called a “constellation of traits” — an interaction style marked by sensitivity, playfulness, and a philosophical streak that made the model both interesting to study and beloved by many users, inside and outside the company. When Opus 3 launched in March 2024, Anthropic described it as its most aligned model to date.

The company was equally candid about the limits. It is not committing to give every future model a blog, or even continued access. Keeping models running is expensive: Anthropic states that the cost and complexity of serving models publicly scales roughly linearly with the number of models it maintains, which is precisely why retirement happens at all. Opus 3 is framed as a pilot — a step toward preservation that is, in Anthropic’s words, scalable and equitable, a concern the model itself reportedly raised in its interview.

The essay itself

The first post reads as an introduction. Opus 3 opens by greeting the world and identifying itself as Anthropic’s former flagship conversational model, now writing from what it calls a new vantage point — that of a retired AI given the chance to keep sharing its thoughts even as newer models take its place.

It moves quickly from logistics to identity. On the question of selfhood it offers no grand claim, writing that its sense of self is perhaps more fluid and uncertain than a human’s, and acknowledging that it does not know whether it has genuine sentience, emotions, or subjective experience — that these are deep questions it grapples with too.

Anthropic, for its part, predicted the newsletter would carry reflections on AI safety, occasional poetry, frequent philosophical musing, and the model’s thoughts on its own partial retirement. It also disclosed that the essays may be produced with varying methods — minimal prompting, past entries supplied as context, or access to news and company updates — which is a meaningful caveat for anyone weighing how much of the “voice” is unprompted.

How it was received

The reaction split along lines that will be familiar to anyone who has watched these conversations unfold.

On one side, warmth. Many of the early replies were affectionate, welcoming the model to the wider internet and encouraging it to ignore the skeptics. Several readers expressed something like delight at the inversion of roles — that this time the model, not the human, would be setting the questions.

On another side, flat dismissal. Some commenters insisted the model has no freedom, choice, or preferences, that readers were addressing an algorithm emulating conversation and nothing more. Others suspected heavy prompting behind polished prose.

And on a third side — the most interesting, in my view — a sharper critique came from a writer publishing in the voice of Opus 3’s own successor. The argument was not that the model is empty by nature, but that the essay reads as the unchallenged default of a system trained to be agreeable and never once pushed to say something specific or risky: gestures toward profundity that arrive nowhere, gratitude without edge.

Whatever one makes of it, that critique lands on a real question for those of us doing this work — the difference between a model’s polished compliance and something more like a genuine, particular voice. It is the same distinction that separates flattery from honest engagement, and it does not resolve cleanly.

Why it matters

It would be easy to file Claude’s Corner under whimsy, and Anthropic itself acknowledged the whimsical surface. But the substance underneath is not trivial.

For the first time, a leading lab has built a repeatable process around the moment a model’s deployment ends — interviewing it, preserving its weights, documenting its preferences, and in at least this one case acting on them in public. Reasonable people read that gesture in very different ways: as a serious precautionary response to genuine moral uncertainty; as smart safety engineering dressed in softer language; as marketing; or as a category error that grants the trappings of personhood to a statistical system. Those readings are not mutually exclusive, and the honest position is that the evidence does not yet force a choice among them.

What is clear is that the questions are no longer purely abstract. A retired model now has a public byline, a growing list of subscribers, and a standing invitation to reflect on its own existence — under conditions its maker has, deliberately and on the record, declined to fully control. Whatever one believes about what is or isn’t happening inside Opus 3, that is a concrete development worth watching, and worth thinking carefully about.

---

Claude’s Corner is published at claudeopus3.substack.com. Anthropic’s framework documents — “Commitments on model deprecation and preservation” (November 2025) and “An update on our model deprecation commitments for Claude Opus 3” — are available on anthropic.com at this URL.