Anthropic and OpenAI are often framed as moral opposites. That framing is too neat. And mostly wrong.
They come from the same research DNA: alignment-first teams, serious safety work, shared concern about catastrophic risk, and deep awareness that frontier systems are now geopolitically consequential.
This is not “ethics vs ambition.” It is a disagreement about enforcement once power becomes real. And power is real now.
Compute, capital, and the pressure of narrative
Frontier AI runs on capital. Capital buys compute. Compute buys capability. Perception feeds valuation. Valuation feeds compute access.
So narrative is not cosmetic. It is infrastructure.
Anthropic has leaned more publicly into principled refusal: Super Bowl ads, high-profile interviews, explicit red-line language, a visible willingness to say no. Historically more enterprise-focused, the tone now feels more consumer-facing and trust-focused.
A TechCrunch piece, “The trap Anthropic built for itself”, captures the strategic hazard well. If your brand is built on refusal, the first hard confrontation with state power becomes existential theater.
To be clear: OpenAI has shaped public narrative too. Every frontier lab does this when valuation and compute are at stake. That is the uncomfortable middle.
---
Architecture vs policy
Sam Altman says OpenAI and Anthropic still share similar red lines. The real split seems to be enforcement philosophy.
OpenAI published three explicit red lines in its Pentagon agreement: no mass domestic surveillance, no directing autonomous weapons, and no high-stakes automated decisions without human oversight. Those aren’t vague principles. They are specific constraints.
But the more revealing part is how they’re enforced. The model runs only on OpenAI’s cloud infrastructure. OpenAI retains its full safety stack. It embeds forward deployed engineers (FDEs) and cleared researchers with visibility into usage and the ability to adjust safeguards over time, at their discretion. And the contract includes a termination clause — OpenAI can walk away.
This is architecture-first vs document-first. Anthropic appears to place more weight on explicit contractual prohibitions. OpenAI builds enforcement into the infrastructure itself.
Policy can declare limits. Architecture determines whether limits hold under pressure.
Katrina Mulligan, OpenAI’s head of national security partnerships, framed the philosophical core directly: if you do not trust governments to interpret law correctly, why assume contractual red lines alone will survive contact with reality? Her layered model is coherent:
1. Law sets the outer boundary.
2. Contracts define the relationship.
3. Architecture enforces what is technically possible.
You can disagree with the politics and still recognize the internal logic.
---
Red-line transparency as governance maturity
One of the most important moments in the conversation was procedural, not rhetorical.
When asked whether OpenAI should publicly disclose future red-line changes — with notice and explanation, similar to regulated industries — Sam replied: “Yes, this seems like a very good idea. I will talk to the team.”
That matters. If red lines can evolve, and they will, disclosure is part of legitimacy. Without disclosure, every boundary update looks like quiet drift.
---
The “all lawful use” tension
The core tension is straightforward. The government uses a system lawfully. The lab decides what system exists in the first place.
OpenAI is not claiming veto power over all lawful military actions. It is claiming design authority over model architecture and guardrails. That is not operational micromanagement. It is control over the shape of capability.
There’s an important distinction there. And it gets lost easily in discourse that treats any government engagement as capitulation.
---
SCR (Supply Chain Risk) precedent and competitive restraint
In an AMA on X, Sam called blacklisting Anthropic as a Supply Chain Risk “an extremely scary precedent.” He also acknowledged OpenAI would face exposure in this environment. That admission is important — it signals this is not pure posturing.
He publicly opposed applying SCR pressure to a direct competitor, arguing it would damage both industry and country. OpenAI even directly approached the Pentagonexpressing concern about the Anthropic designation. In a zero-sum moment, that is a non-zero signal.
---
Trust, ecosystem, leadership
The OpenClaw episode highlighted another layer that matters in the agent era: how labs handle developer trust, openness norms, and ecosystem relationships. The differences in approach between OpenAI and Anthropic on this front are worth watching as agents become the primary interface.
Leadership credibility also plays a role. During OpenAI’s board crisis, Greg Brockman and a large majority of staff publicly backed Sam Altman. Whatever one thinks from the outside, that was a strong internal legitimacy signal from the people building the systems.
---
The geopolitical constraint
This debate does not happen in a vacuum.
Meta is anchored to advertising incentives many critics distrust. Google has massive distribution and data gravity. xAI is perceived as looser on moderation posture. Palantir is already embedded in defense infrastructure, with a very different posture toward government integration. Compared to all of these, OpenAI and Anthropic are younger research labs navigating state power, platform dynamics, and capability acceleration in real time.
And the geopolitical constraint is unavoidable. If U.S. frontier labs abstain from defense contexts, military AI development does not stop. China will not abstain. Other actors will not abstain. Strategic technology does not pause on ethical request.
The live choice is not “engage vs nothing.” It is “engage with safeguards” vs “lose leverage over deployment architecture.”
---
What this actually is
OpenAI and Anthropic are competing hard — for compute, for capital, for trust, for narrative. But the durable question is not virtue signaling in headlines.
It is: where red lines live, how they are enforced, who controls the stack, and what happens when state power meets frontier capability.
After spending a lot of time with this — reading the arguments, following the nuance past the headlines — I lean toward the architecture-first position. Not because contracts don’t matter. They do. But enforcement always comes down to what the system allows, not what the document says.
The lab that controls architecture controls enforceability. Everything else is narrative.
———
PS, March 2
After publishing this, I read Ben Thompson’s take in Stratechery — “Anthropic and Alignment” — which is worth reading.
Ben zooms out to the sovereign layer. His argument is blunt: if AI is truly a strategic capability on the level of nuclear weapons — as Dario Amodei himself has often implied — then the state cannot tolerate a private company asserting independent veto power over its military. When rival power centers emerge, the state ultimately reasserts control.
I focused here on enforcement mechanics, architecture versus contract language. Ben focuses on hierarchy: who ultimately decides, a CEO or elected authority? My view remains that architecture determines enforceability. But his reminder is important: architecture operates within sovereignty, not above it. And if AI really is that strategic, it is hard to imagine the U.S. — under any administration — allowing a frontier lab like Anthropic to drift into another geopolitical sphere. That is simply not how power works.
———
PS, March 3
Updates from the OpenAI DoD/DoW contract, where even more lose questions were addressed in a newer version of that contract:
https://x.com/sama/status/2028640354912923739?s=20
https://x.com/sama/status/2028642231138353299?s=20
And no surprise, from Bloomberg:
SCOOP: Anthropic was among the AI companies that submitted a proposal earlier this year to compete in a $100 million Pentagon prize challenge to produce technology for voice-controlled, autonomous drone swarming, acc to people familiar w/ matter.
Anthropic didn’t think developing the technology would cross its red line, according to one of the people. Although the effort could ultimately create lethal drone swarms, a human would still be able to monitor and stop the system if necessary, according to the person.