Anthropic walked away from the table. OpenAI stayed and negotiated. In the long run, engagement—not abstention—may prove the more responsible path.

In the final week of February 2026, the American AI industry confronted a question it had been avoiding for years: What happens when the world’s most powerful technology meets the world’s most powerful military, and the two cannot agree on the rules? The answer arrived in dramatic fashion. Anthropic, maker of the Claude AI assistant, refused to grant the Pentagon unrestricted access to its models, insisting on contractual prohibitions against mass domestic surveillance and fully autonomous weapons. Within hours of that refusal, President Trump ordered all federal agencies to cease using Anthropic’s products. Defense Secretary Pete Hegseth designated the company a “supply chain risk to national security”—a label never before applied to an American firm.

OpenAI moved into the vacuum. CEO Sam Altman announced a deal with the Department of Defense that he said preserved the same red lines Anthropic had demanded—prohibitions on domestic surveillance, autonomous weapons, and high-stakes automated decision-making—while providing the military access to frontier AI on classified networks.

The public narrative quickly calcified into a morality play: Anthropic, the principled martyr; OpenAI, the opportunistic profiteer. Claude surged to number one on the App Store. ChatGPT uninstalls spiked nearly 300%. Anthropic’s CEO Dario Amodei called OpenAI’s messaging “straight up lies” and “safety theater.”

But the morality play obscures a more complicated—and arguably more important—strategic reality. In the long arc of AI governance, OpenAI’s decision to stay at the table and negotiate may prove to be the more consequential act of responsibility.

The Case for Engagement

The most important fact about the OpenAI-Pentagon deal is not what it permits, but what it structurally prevents. OpenAI’s models are deployed exclusively through the company’s cloud infrastructure. They are not installed on edge devices—no drones, no fire-control systems, no autonomous platforms. The company retains full discretion over its safety stack, meaning it controls what the models will and will not do at the technical level. Cleared OpenAI engineers and safety researchers remain “in the loop” for sensitive workflows.

This is not a blank check. It is a deployment architecture designed to make misuse structurally difficult. If the Pentagon attempts to wire OpenAI’s models into an autonomous weapons system, the cloud-only design makes that functionally impossible without OpenAI’s active cooperation. If the government attempts to repurpose the tools for mass surveillance, OpenAI’s personnel embedded in the process would have visibility into those attempts, and the contract gives OpenAI the right to terminate. This is a strategy of proactive participation over reactive restriction. OpenAI is taking a seat at the table to co-create the norms and rules for AI in the domain of war, a space that Anthropic’s silence and distance leaves dangerously open. The only way to ensure that AI is used responsibly in defense is for the responsible creators of that technology to be actively involved in that process.

Anthropic’s original contract, by contrast, was deployed through Palantir and had already been running on classified networks for months. When Anthropic asked Palantir whether its models had been used in the military’s raid on Venezuela that captured Nicolas Maduro, the question itself triggered alarm bells at the Pentagon—suggesting Anthropic had limited real-time visibility into how its tools were actually being used in the field.

The irony is sharp. Anthropic’s insistence on contractual red lines may have been a response to the fact that its existing deployment model gave it relatively little technical control. OpenAI, entering later, built its deployment around a fundamentally different architecture—one designed to give the company ongoing, structural leverage over how its technology is used.

The Limits of Principled Refusal

Anthropic’s position commands respect. Dario Amodei’s public statement articulated a genuine concern: that AI-driven mass surveillance presents novel risks to civil liberties, and that frontier models are not yet reliable enough to power fully autonomous weapons. Both points are defensible. Both deserve serious engagement from policymakers.

But Anthropic’s stance also reveals a tension at the heart of principled refusal. By walking away, Anthropic lost its seat at the table. It lost visibility into how AI is being used in active military operations. It lost the ability to shape norms from the inside. And it created a vacuum that was immediately filled—not only by OpenAI, but by Elon Musk’s xAI, which agreed to deploy its models across classified systems as well.

Retired General Paul Nakasone, the former director of the National Security Agency and now an OpenAI board member, captured the practical reality at an Aspen Institute event: “We need Anthropic, we need OpenAI, we need all of our large language model companies to be partnering with our government.” His point was not about patriotism—it was about redundancy and competition. A healthy AI-defense ecosystem requires multiple providers with different approaches, pushing each other and the government toward better practices.

When there is only one company willing to engage, the government’s leverage over that company increases and the company’s leverage over the government diminishes. Paradoxically, Anthropic’s refusal may have weakened the very guardrails it sought to protect, by reducing the competitive pressure that incentivizes governments to accept stricter terms.

The “All Lawful Purposes” Question

The central contention in the Anthropic-Pentagon breakdown was the phrase “all lawful purposes.” The Department of Defense insisted that AI companies accept this standard. Anthropic argued the phrase was too broad—that current law has not caught up with AI’s capabilities, and that conduct technically legal today (such as purchasing Americans’ movement and browsing data from commercial brokers without a warrant) could enable surveillance that is incompatible with democratic values.

This is an important argument, but it suffers from a structural problem: it asks a private company to substitute its judgment for that of democratically elected officials and the courts on questions of constitutional scope. There is a difference between a company saying “we won’t build tools designed for mass surveillance” and a company saying “we, not Congress, will decide what constitutes acceptable surveillance under American law.”

OpenAI’s approach threads this needle differently. Rather than demanding the government accept restrictions beyond what the law requires, OpenAI references specific existing legal authorities—the Fourth Amendment, the Foreign Intelligence Surveillance Act, Executive Order 12333, DoD Directive 3000.09—and contractually binds the government to those standards even if future administrations attempt to loosen them. The contract further states that the AI system “shall not be intentionally used for domestic surveillance of U.S. persons and nationals” and shall not “independently direct autonomous weapons” where policy requires human control.

This is not a perfect framework. Critics have rightly noted that the word “intentionally” creates ambiguity, and that Executive Order 12333 has historically been interpreted to permit collection of Americans’ data when captured incidentally through overseas surveillance. But it is a framework anchored in the legal system Americans have built to govern their military—not in the unilateral judgment of a Silicon Valley CEO.

•  •  •

The De-Escalation Argument

There is a dimension of OpenAI’s decision that has received insufficient attention: the company explicitly asked the Pentagon to offer the same terms to all AI labs, including Anthropic. Altman publicly stated that Anthropic should not be designated a supply chain risk. OpenAI’s blog post expressed hope that Anthropic and other companies would accept the deal framework.

This is not the behavior of a company seeking to exploit a rival’s misfortune. It is the behavior of a company that recognized the Anthropic-Pentagon standoff was spiraling toward a catastrophic precedent—one in which the government could effectively destroy any AI company that resisted its demands—and attempted to create an off-ramp.

Altman acknowledged the move was “definitely rushed” and that “the optics don’t look good.” At an all-hands meeting, he told employees, “If we are right and this does lead to a de-escalation between the DoW and the industry, we will look like geniuses. If not, we will continue to be characterized as rushed and uncareful.”

The gamble is worth understanding in its full context. Had no major AI company accepted the Pentagon’s terms, the government had explicitly threatened to invoke the Defense Production Act—a wartime power that would allow it to compel AI companies to provide technology without any negotiated safeguards at all. The choice was not between a perfect contract and an imperfect one. It was between an imperfect contract with real guardrails and the possibility of no contract and no guardrails.

The Uncomfortable Middle Ground

None of this means OpenAI’s deal is above criticism. The contract has not been publicly released. The cloud-only restriction, while significant, is not an absolute barrier—advanced militaries increasingly operate through cloud-connected mesh networks that blur the distinction between cloud and edge. But they also exist within a broader strategic context that the morality-play narrative ignores.

The United States is in an active conflict. AI tools are already being used by the military for intelligence analysis, operational planning, and target identification in the ongoing operations against Iran. Pentagon Under Secretary Emil Michael has publicly described how deeply embedded Anthropic’s own models had already become in defense operations—and how alarmed officials were at the prospect of losing access during active combat.

In this environment, the question is not whether AI will be used for defense. It will be. The question is whether the companies building that AI will be at the table, shaping how it is used, maintaining technical control over their systems, and retaining the contractual right to walk away if red lines are crossed—or whether they will cede that ground to less scrupulous actors, foreign and domestic.

What Comes Next

The OpenAI-Anthropic split may prove to be the most consequential corporate divergence in the brief history of the AI industry. But its significance lies not in who was “right” and who was “wrong,” but in what it reveals about the two available models for governing AI in national security: engagement with imperfect guardrails, or refusal with no guardrails at all.

History suggests that engagement, while messier, tends to produce better outcomes. The defense industrial base that built America’s nuclear arsenal also produced the norms and institutions—arms control treaties, civilian oversight, the laws of armed conflict—that have kept those weapons from being used since 1945. Those norms were not established by companies that refused to participate. They were built by people who stayed in the room.

Anthropic’s instinct to draw bright lines around mass surveillance and autonomous weapons is correct. But bright lines on paper are worth less than structural controls in practice. OpenAI’s cloud-only deployment, in-the-loop personnel, retained safety stack, and contractual termination rights are not guarantees—nothing is. But they are the kinds of practical, enforceable mechanisms that can evolve into genuine governance norms as the technology matures.

The applause for Anthropic’s refusal is understandable. It is always easier to cheer the company that says no. But in a world where AI will be used for defense whether Silicon Valley likes it or not, the harder and more important question is not whether to participate, but how to participate responsibly. On that question, OpenAI’s answer—imperfect, rushed, politically costly—deserves more credit than it has received. OpenAI’s decision to engage with the “Department of War” situation is a masterstroke of geopolitical pragmatism. In the new reality of AI-enabled statecraft, it is the only viable path forward. Here is why OpenAI’s stance is the correct one for a future of stable AI governance.

The Geopolitical Mandate for Pragmatism

The single most important factor is the inescapable reality of global competition. Advanced AI is a strategic asset. If Western democracies, guided by well-intentioned but overly restrictive policies, refuse to utilize the capabilities of their best AI models (like GPT-4) in their defense architecture, they will be ceding that terrain. Adversarial nations, like China and Russia, face no such moral qualms. They are aggressively investing in the integration of AI into every facet of their militaries, from cyberwarfare to autonomous systems.

A world where authoritarian states possess vastly superior, unrestricted AI capabilities is a scenario for a global security catastrophe. OpenAI’s decision acknowledges this: if the U.S. and its allies do not have access to the leading-edge technology developed on their own soil, the resulting capability gap would be a gift to their adversaries. The vacuum of a purely restrictive policy would not be filled with peace, but with a more dangerous and unstable arms race of autonomy.

OpenAI is not creating “terminators”; it is providing the intelligence for cybersecurity defenses, logistics management, medical analysis, and intelligence processing. It is using its sophisticated models to strengthen the backbone of a responsible democratic state’s defensive power. The choice is not between safe AI and dangerous military AI, but between US-developed, safety-conscious systems being used by responsible actors versus foreign, unrestricted, and likely less safe systems dominating the global arena.

In the nascent and volatile battle for AI governance, a spectrum of stances is inevitable and healthy. Anthropic’s focus on pure-form, theoretical safety provides an essential counterweight. But in a world of complex realpolitik and immediate national security challenges, it is OpenAI’s pragmatic, engaged approach that offers the only sound blueprint for a secure future. By stepping into the fray of Department of War engagement, OpenAI is not compromising its ethics; it is ensuring that its technology is part of a global architecture that prioritizes responsible governance and the security of democratic ideals.