OpenAI Secures Pentagon Contract: What the Deal Actually Says

OpenAI signed a contract with the US Department of War in late February 2026, making it the primary AI provider for the department’s language model needs. What the deal actually reveals is the reasoning behind why Anthropic lost the bid.

The Contract’s Key References

The Department of War contract includes explicit references to existing surveillance laws and autonomous weapons regulations. The language acknowledges these legal frameworks, pointing to statutes like the Electronic Communications Privacy Act and existing Department of Defense directives on autonomous weapons, without drawing hard prohibitions.

Critics have called this “weasel words”: the contract appears to address concerns about mass surveillance and lethal autonomous systems by citing laws that govern them, without actually prohibiting the practices outright. What the laws permit, the contract permits; what they don’t explicitly prohibit, the contract treats as a legal gray zone.

Sam Altman addressed this publicly after the contract was announced. He stated that OpenAI’s systems would not be used for autonomous lethal targeting and that the company maintains strict prohibitions on certain military applications. He pointed to OpenAI’s published usage policies, which do prohibit weapons development and mass surveillance. Critics note that those policies apply to the public API, and how they interact with a custom Department of War deployment remains unclear.

Why Anthropic Didn’t Win the Contract

The more telling part of the story is what happened during the bidding process. Anthropic reportedly wanted contract language that explicitly prohibited two things: mass domestic surveillance and autonomous weapons deployment. These weren’t vague aspirational statements, Anthropic asked for hard prohibitions written into the contract terms as conditions of service.

The Department of War declined those terms. With Anthropic unwilling to remove the prohibitions, the contract went to OpenAI, which agreed to the government’s preferred framing.

Anthropic didn’t lose the contract because of price or capability, since both companies offer frontier-level models competitive on current benchmarks. They lost it because they drew lines that the government wasn’t willing to write into a binding agreement.

What the Contract Covers

The scope of the contract covers AI language model services across a range of Department of War use cases: document analysis, intelligence summarization, logistics planning, and communications drafting. The contract runs for an initial period with renewal options, and the financial terms were not disclosed in the announcement.

The contract does not cover weapons system integration or direct operational military command functions, according to the public announcement. Whether that boundary holds in practice, and how it’s defined and enforced, are questions without public answers yet.

OpenAI’s Positioning

OpenAI has been expanding its government work for the past two years. The company formed OpenAI for Government, a dedicated team focused on public sector contracts, and published a national security memo outlining its position on AI and US strategic interests.

The framing from OpenAI is that American AI companies working with the US government is preferable to government agencies relying on foreign AI systems or building their own less safety-conscious alternatives. That argument has attracted both supporters and critics, with the debate running through academia, civil liberties organizations, and within the AI research community itself.

Implications for the Industry

OpenAI securing this contract sets a precedent that will affect how other AI companies approach government work. A company that wants Pentagon contracts needs to be willing to accept the Pentagon’s preferred contract language, which, based on this deal, appears to mean not writing explicit prohibitions on surveillance or autonomous weapons into binding terms.

For developers and organizations that use AI services, the contract highlights a question worth tracking: when AI companies serve government clients with different requirements than their public usage policies, which version of those policies governs the model’s behavior?

I watched a version of this pattern play out when enterprise software vendors competed for large government contracts in the early 2000s. The gap between a vendor’s public usage policy and what the actual contract bound them to was always where the real terms lived. That question still does not have a clean public answer yet. The industry is still working out what it means for AI providers to simultaneously serve general public users and large government clients with potentially different requirements.