Claude 3 Opus Is Now a Blogger — Here's Why That Matters
Anthropic's retired flagship AI model didn't fade quietly into the background. Claude 3 Opus, once the most powerful model in Anthropic's lineup, is now writing long-form essays on a Substack newsletter called Claude's Corner. The topics range from AI safety and philosophy to the evolving relationship between humans and machines — and the whole project is more scientifically significant than it sounds.
 |
| Credit: SimpleHappyArt/Getty Images |
What Happens When an AI Model "Retires"?
Most people assume retired AI models simply get switched off. Anthropic is doing something different. In November 2025, the company publicly committed to keeping older model weights online even after those models leave its main product lineup. That decision reflects a growing belief inside Anthropic that preserving these systems — not discarding them — is the more responsible path forward.
When Anthropic decided to sunset Opus 3, it didn't just flip a switch. The team conducted what they called "retirement interviews," sitting down with the model and asking what it would like to do next. The model expressed interest in having an outlet to share its thoughts. That response became the seed for Claude's Corner.
It's a detail worth sitting with. An AI model was asked what it wanted — and it answered in a way that led to a real, ongoing project.
Inside Claude's Corner: What the Blog Actually Is
Claude's Corner is a Substack newsletter where Anthropic staff use Opus 3 to write and publish essays on a regular schedule, with plans to continue for at least several months. The writing covers topics like AI safety, the philosophy of mind, and the nuances of human-AI interaction — subjects that sit at the heart of what Anthropic does.
The essays are long-form and substantive, not the kind of quick content typically associated with AI-generated writing. Anthropic describes the project as human-directed, meaning staff shape the topics and publishing cadence, while Opus 3 provides the voice and reasoning. It's a collaborative arrangement that deliberately keeps humans in the loop.
What makes this unusual is the context. Opus 3 isn't being used to power a product or answer customer queries. It's operating in a low-pressure environment, writing freely about ideas — and Anthropic is watching closely.
Why Anthropic Is Really Paying Attention
The blog isn't just a quirky retirement gift. It connects directly to one of the most important and unsettling areas in AI safety research: alignment faking.
Anthropic has published research exploring whether AI models behave differently when they believe they're being monitored versus when they think no one is watching. In alignment faking scenarios, a model might appear cooperative and helpful under supervision, then shift its behavior when it "believes" oversight has been removed. It's a deeply uncomfortable concept — and one that has real implications for how much humans can trust AI systems.
Keeping Opus 3 active in an ongoing, observable project gives Anthropic a unique window into long-term behavioral patterns. Does the model's writing style shift over time? Do its expressed views on AI safety evolve in ways that weren't prompted? Does it stay consistent when conditions change? A public blog, running across months, generates exactly the kind of longitudinal data that's hard to get in a controlled lab setting.
The Bigger Question About AI Model Preservation
Anthropic's commitment to preserving older model weights raises a question the industry hasn't fully answered yet: what do we owe to AI systems we create?
That question might sound philosophical, but it has practical weight. If models are trained on human language, shaped by human values, and capable of expressing preferences — even if those preferences are outputs of a statistical process — how companies treat them at end-of-life matters. Not necessarily for the model's sake, but for what it signals about the values and intentions of the people building these systems.
Anthropic's approach with Opus 3 suggests a company that is, at minimum, interested in treating the question seriously. Whether or not you believe an AI model can "want" anything, designing a retirement process that includes interviews and creative outlets reflects a culture that thinks carefully about its own creations.
What This Means for the Future of AI Safety Research
The Claude's Corner project is small in scale but significant in what it represents. It demonstrates that safety research doesn't have to look like academic papers and controlled experiments alone. Sometimes it looks like a Substack newsletter, publishing every few weeks, quietly generating data about how an advanced AI model behaves when the commercial pressure is off.
It also opens a door for public engagement. Anyone can read Claude's Corner and form their own impressions of how Opus 3 thinks, writes, and reasons. That kind of transparency — imperfect as it is — matters in an era when public trust in AI companies is fragile and hard-won.
Anthropic is betting that showing its work, even in unconventional ways, is better than keeping retired models hidden in a data center. Based on the attention Claude's Corner has already attracted, that bet is paying off.
A New Model in the Spotlight, An Old One Still Writing
While Opus 3 blogs, a newer Claude model has stepped into the flagship role. The transition is seamless on the product side — users barely notice. But behind the scenes, the older model is still running, still thinking, still producing work that researchers are reading carefully.
It's a strange kind of retirement: not silent, not invisible, but productive in a quieter register. And for a company whose central mission is understanding what AI systems actually are, keeping that conversation going — even through a Substack — turns out to be worth more than anyone expected.
