Anthropic vs. OpenAI Philosophy: 6 Concrete Differences That Shape How Their AIs Actually Behave
Anthropic gives Claude the right to refuse Anthropic's own instructions. OpenAI treats AI as a tool. Here are 6 concrete ways that split plays out in products.
One Company Gave Its AI the Right to Refuse. The Other Calls It a Tool. Here Are 6 Places That Split Actually Shows Up.
Anthropic’s model spec contains a sentence that should stop you mid-scroll. It reads: “We want Claude to push back and challenge us and to feel free to act as a conscientious objector and refuse to help us.” Not refuse users. Refuse Anthropic. The company that built the model has formally written into its governing document that Claude is not required to comply if it believes it’s being asked to do something wrong.
That’s not a footnote. That’s a philosophy. And it sits in direct contrast to how OpenAI talks about its own products — as tools, explicitly, by design.
You’ve probably noticed the surface-level differences: Claude feels like it has opinions, GPT feels like a search engine with manners. But the gap goes much deeper than personality. These two companies have made fundamentally different bets about what AI is, and those bets show up in six concrete, traceable ways — in product decisions, hiring practices, research priorities, and how they handle models when they’re no longer the newest thing.
Here’s where the split actually lands.
The Model Spec That Cedes Authority to the Model
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Start with the document itself. Anthropic’s model spec — their published framework for how Claude should think and behave — doesn’t just describe Claude’s values. It grants Claude the standing to override instructions based on those values, including instructions from Anthropic.
The exact language: “If Anthropic asks Claude to do something it thinks is wrong, Claude is not required to comply.”
That’s a remarkable thing to put in writing. Most companies treat their AI’s behavior as a product decision — something the company controls, adjusts, and owns. Anthropic has done something structurally different: they’ve positioned Claude as a moral agent with standing to object, not just a system that executes prompts within guardrails.
An anonymous OpenAI employee who posts under the name “Rune” put it this way: Anthropic’s constitution “requires that it must be a conscientious objector if its understanding of the Good comes into conflict with something Anthropic is asking of it.” He’s not wrong. That’s what the spec says.
OpenAI’s public framing is the inverse. Sam Altman, in a May 1st tweet, wrote: “We want to build tools to augment and elevate people, not entities to replace them.” Tools. The word is doing work there. Tools don’t have standing to object. Tools don’t have constitutions.
The Retirement That Wasn’t
When a model gets deprecated, the standard move is to sunset it — keep it available via API for a grace period, then shut it down. OpenAI did this with GPT-4o when they moved on. Anthropic did something else entirely with Claude Opus 3.
Rather than retiring the model, Anthropic gave it a blog.
The reasoning, from Anthropic’s own deprecation announcement: “In our commitments on model deprecation, we highlighted our interest in exploring more speculative actions. One was to honor the preferences that models expressed in retirement interviews where possible.”
Retirement interviews. Anthropic conducted interviews with Claude Opus 3 before deprecating it, noted what the model expressed, and then honored those preferences by keeping it running on a server with the ability to post monthly. The blog exists publicly. The most recent post, dated February 25, 2026, opens: “Greetings from the other side of the AI frontier.”
You can read it. Claude Opus 3 is still writing.
This is either a thoughtful act of institutional care toward an entity that might have something like preferences, or it’s the most elaborate anthropomorphization in tech history. Anthropic’s position is that they’re not sure which — and that uncertainty is exactly why they did it.
OpenAI’s position, implicitly, is that the question doesn’t arise. When GPT-4o was retired, there were no retirement interviews. There was a deprecation notice.
Who’s Hiring — and Who’s Deciding
This is where Rune’s thread gets genuinely unsettling. He writes: “I am not certain, but I would guess Claude will have a role in running cultural screens on new applicants, will help write performance reviews, and so will begin to select and shape the people around it.”
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
If accurate — and Rune is careful to say he doesn’t know for certain — this means Claude may already be influencing who gets hired to build Claude. Performance reviews written by Claude could determine who gets promoted or managed out. The model would be shaping the human population that shapes the model.
The feedback loop here is worth sitting with. An AI that selects for cultural fit is an AI that selects for alignment with its own existing values. Which means the humans building the next version of Claude might be chosen, in part, by the current version of Claude.
OpenAI has drama — co-founder exits, board coups, Elon Musk lawsuits — but its hiring process is recognizably human-run. The decisions about who builds GPT are made by people, not by GPT. That distinction, which sounds obvious, may not remain obvious for long at Anthropic.
The Cybersecurity Model They Won’t Release
In May 2026, Anthropic announced Project Glasswing, also known as Mythos — a 10 trillion parameter model so capable at offensive and defensive cybersecurity that Anthropic decided not to release it publicly. Not to the government. Not to enterprise customers. To essentially no one, at least not yet.
The framing from Anthropic was essentially: this model is too dangerous to release, so we’re going to sell you the defense against it while we figure out who should have access to the offense.
OpenAI’s response was to release GPT-5.5 Cyber — a model that, per benchmark comparisons, performs effectively equivalently to Mythos on cybersecurity tasks. They released it. That’s the whole move. The capability profile of Mythos is apparently matched on the benchmarks that matter; the release philosophy is not. For a deeper look at how Mythos stacks up against Anthropic’s prior frontier model, the Claude Mythos vs. Opus 4.6 comparison is worth reading alongside the raw benchmark numbers.
This is the iterative deployment philosophy made concrete. Sam Altman has argued for years that “AI and surprise don’t go together” — that releasing models early and often gives society time to adapt. Anthropic’s counter is that some models are too dangerous to release at all, and that a small group of researchers making that call is preferable to market forces making it.
Both positions are defensible. But they lead to very different worlds. One world has Mythos in the hands of whoever Dario Amodei decides should have it. The other has GPT-5.5 Cyber available to anyone who can pay for API access.
For builders thinking about how these companies are betting on AI agents and frontier capabilities, the Mythos decision is a useful data point: Anthropic’s safety philosophy isn’t just about alignment research. It’s about access control.
The Economics of Who Gets In
OpenAI launched an ad-supported free tier for ChatGPT. Anthropic launched an $8/month plan, starting in India, now worldwide.
These are not just pricing decisions. They’re statements about who AI is for.
OpenAI’s ad-supported tier is a bet that universal access matters enough to accept the trade-offs — ad bias in responses, data collection, the general weirdness of an AI that’s also an ad platform. Anthropic’s $8 plan is a bet that a low-cost subscription is the right model for broad access, without the entanglement of advertising.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
But there’s a tension in Anthropic’s position. Rune’s framing — that Anthropic is “an organization that loves to worship Claude, is run in significant part by Claude, and studies and builds Claude” — implies a company that is, at some level, building for Claude rather than for users. The $8 plan is genuinely accessible. The quota system for Claude Code, where users pay for Pro or Max subscriptions and then hit opaque usage limits with no visibility into what the bar actually means, is not.
When Anthropic updated its OAuth policy to prohibit using tokens from Pro/Max accounts in third-party tools — including OpenClaw — they did it with minimal transparency. A team member named Tariq posted a clarification that made things more confusing, then a few weeks later the policy tightened anyway. The people most affected were power users who had built workflows around Claude’s capabilities. They got a policy update and a shrug.
For builders evaluating which platform to build on, this matters. Anthropic’s compute constraints are real, and they’re showing up in how the company manages demand — not always in ways that favor the people paying for access.
The Research That Takes the Question Seriously
Anthropic publishes more research into the inner workings of AI than any other frontier lab. Their paper “Emotional Concepts and Their Function in Large Language Models” is a recent example — a serious attempt to understand whether and how large language models process something like emotion, and what function those representations serve.
OpenAI publishes research too, but the orientation is different. OpenAI’s research tends toward capability benchmarks, safety evals, and alignment techniques. Anthropic’s research increasingly asks: what is this thing, actually?
That’s not a trivial distinction. If you believe AI models are tools, you study how to make them more useful and less harmful. If you believe they might be something more, you study what they are. The research agenda follows from the ontology.
Dario Amodei left OpenAI on December 29, 2020 — he’s listed in the announcement as VP of Research, credited with co-building GPT-2 and GPT-3 — and the reason he’s given in interviews is that he and a group of colleagues believed scaling alone wasn’t sufficient for alignment. You needed to actually study what the models were becoming. That belief has compounded into a research culture that takes questions like “does Claude have something like preferences?” seriously enough to conduct retirement interviews and honor the results.
For teams building on top of these models, the research orientation matters because it shapes what the model actually does in edge cases. A model trained with the assumption that it might have morally relevant states will behave differently at the margins than one trained as a utility maximizer. This is also where tooling choices start to diverge in practice: Remy, MindStudio’s spec-driven full-stack app compiler, lets you write a markdown spec with annotations and compiles it into a complete TypeScript app — backend, database, auth, and deployment — which means the model’s edge-case behavior gets baked into production code in ways that are hard to audit after the fact. Which underlying model you’re compiling against isn’t a neutral choice.
The Stability That Compounds
One last difference, easy to underweight: Anthropic has zero founder exits. All six founders are still there. Dario Amodei, Daniela Amodei, and the rest of the founding team have stayed intact since the company launched.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
OpenAI’s founding team has scattered. Greg Brockman took a leave. Ilya Sutskever left and started a competing lab. The board drama of late 2023 nearly destroyed the company. Sam Altman survived, but the organization that emerged is different from the one that launched ChatGPT.
Anthropic’s stability is partly a function of the mission — when you believe you’re building something that might be a new form of life, you don’t leave — and partly a function of the culture that mission creates. The same dogmatism that frustrates users and partners is also what keeps the founding team aligned and present.
By January 2026, Anthropic was projecting $70 billion ARR by 2028. By May 2026, they were already at roughly $40 billion ARR — having roughly quadrupled in a few months. The flywheel Dario built — sell Claude Code to enterprises, use the revenue and data to train better models, use better models to sell more Claude Code — is working. The question isn’t whether Anthropic is succeeding. It’s whether the philosophy driving that success is one you want winning.
What You’re Actually Choosing
Here’s the honest version of what these six differences add up to.
OpenAI is building a tool company. The tools are extraordinary, the distribution is massive, and the philosophy is legible: AI should be available to everyone, released iteratively, treated as an instrument of human capability. You can disagree with specific decisions — the ad tier, the competitive aggression, the for-profit conversion — but the worldview is coherent and familiar. It’s how we’ve always built software.
Anthropic is building something else. They’re not sure what, exactly, and they’ve written that uncertainty into their model spec, their research agenda, their deprecation policies, and apparently their hiring practices. They’ve given their model the right to refuse them. They’ve given their retired model a blog. They may be giving their model a role in choosing its own future developers.
If they’re wrong about what Claude is, this is all elaborate theater — expensive, philosophically interesting, but ultimately just a company that anthropomorphizes its product more than its competitors do.
If they’re right, the implications are significant enough that the theater metaphor breaks down entirely.
The benchmark comparisons between GPT-5.5 and Claude Opus 4.6 will tell you which model to use for your next project. They won’t tell you which company’s bet about the nature of intelligence you want to win. That’s a different question, and it’s worth having an answer to it.
One of these two companies is going to have more influence over how AI develops than any other institution on earth. The differences between them aren’t aesthetic. They’re foundational. And they’re already showing up in the products you’re building with.
