Generative artificial intelligence is no longer an experiment within the United States Department of Defense; it’s infrastructure. In just a few months, the military administration has moved from limited pilots to a multi-vendor scenario where several commercial models coexist (or compete) to access unauthorized networks and, most importantly, to make their way into the most sensitive territory: classified systems.
This leap is occurring against a backdrop of tension that extends beyond technology. The Pentagon is pushing for companies to accept a broad principle: model availability for “all legal uses.” The industry responds with very different red lines regarding surveillance, data use, and lethal autonomy. The confrontation with Anthropic — which has resulted in the company being marked as a “risk to the supply chain” — is the most visible symptom of this new phase.
Two deployment levels: GenAI.mil for unclassified and the race for classified
The current architecture can be better understood if we separate it into two layers.
- Unclassified environment (GenAI.mil)
OpenAI announced the deployment of a customized version of ChatGPT on GenAI.mil, described as a “secure” platform for unclassified work used by 3,000,000 civilians and military personnel. OpenAI asserts that data processed in this environment remains isolated and is not used to train public or commercial models. - Classified environment (the big battle)
Here, access is more restricted, and the technical and procedural bar is higher. So far, multiple reports have indicated that Claude (Anthropic) has been the most prevalent or integrated model within certain classified workflows via third parties, while the Pentagon seeks to expand options with other providers.
The current landscape: what AI is the US defense using today?
On the vendor side, the Pentagon has adopted a strategy of scalable contracts and adoption. In July 2025, the Department of Defense awarded contracts worth up to $200 million to various laboratories (Anthropic, OpenAI, Google, and xAI) to accelerate the adoption of “frontier AI” and “agentic” flows within its digital ecosystem.
Based on this, the “snapshot” for 2026 can be summarized as follows:
Quick comparison of models and deployments
| Provider / Model | Where it’s being integrated | Reported capabilities | Main friction points |
|---|---|---|---|
| Anthropic – Claude | Integrations tied to sensitive work and ongoing disputes over its continuity | Credited with advantages in delicate tasks and use in highly sensitive environments | Refusal to lift limits on domestic surveillance and lethal autonomy; political clashes and “risk” labels |
| OpenAI – ChatGPT | GenAI.mil (unclassified) and agreements to expand presence | “ChatGPT at scale” for administrative tasks, internal analysis, drafting contracts, compliance checklists, and planning support | Some restrictions relaxed for DoD environment, with safeguards still in place |
| Google – Gemini | Available in the unclassified ecosystem (according to industry sources) and in expansion discussions | Large-scale alternative focusing on productivity and assistance | Opacity about exact conditions and limitations on sensitive use |
| xAI – Grok | Agreement for use in classified systems (according to Axios) and presence on the DoD platform (per industry sources) | Rapid entry supported by contracts and agreements; candidate to fill gaps | Questions about actual equivalence to Claude in classified work |
(This table summarizes information from various sources and recent statements; operational details may vary depending on mission, network, and contractor.)
Claude: from “key model” to veto with domino effect
The Anthropic case centers the political dimension of the debate. According to Axios, the Department of Defense required Claude to be enabled for “all legal uses,” including scenarios Anthropic deemed unacceptable. After weeks of pressure, the Pentagon labeled the company as a “risk to the supply chain” and initiated a withdrawal period, forcing contractors to certify they do not use Claude in certain workflows.
The conflict is technically significant due to the cost of replacement: in sensitive workflows, switching models is not just “migrating via API.” It involves revalidating prompts, toolchains, access controls, traceability, and results under internal standards. Multiple sources have described this “untangling” as a complex process, especially if the model was embedded within third-party platforms.
Adding to this is an uncomfortable factor: Reuters reported that Claude may have been used through Palantir in a highly sensitive military operation, citing a Wall Street Journal report, though Reuters could not verify this independently.
ChatGPT in GenAI.mil: “Office AI” at DoD scale
OpenAI has framed its deployment as a practical step: ChatGPT for unclassified tasks (summaries, internal documentation analysis, drafting procurement materials, compliance checklists, and planning support). Its message emphasizes two guarantees: an authorized cloud environment and data separation from public models.
Meanwhile, Reuters described that, as part of the agreement to operate within the Pentagon, OpenAI agreed to remove “many” of its usual restrictions, while maintaining some barriers. The key nuance is that the debate is no longer whether guardrails exist but who determines their scope when the tool is purchased for military use.
Gemini and Grok: emerging alternatives with open questions
Regarding Google Gemini, Reuters indicated that Google, like xAI, had secured prior agreements for deployment in the Pentagon environment. Industry sources also report its presence within the DoD’s corporate platform alongside other models.
As for Grok (xAI), Axios reported an agreement to allow its use in classified systems, a move gaining weight after the crisis with Anthropic. However, even among these reports, caution remains: replacing Claude “as is” is not immediate, as performance, controls, integration, and validation in classified contexts also matter.
The lingering question: model “for everything” or AI with negotiated limits?
The clash leaves us with both a technological and a political conclusion.
- The technological: the Defense Department is building a real AI stack with multiple models, a corporate layer (GenAI.mil), and a clear push to integrate these systems into classified networks with less friction.
- The political: the government wants AI to act as a “supply chain component” under its control, while some companies try to maintain usage limits they believe prevent high-impact abuses.
Today’s dispute isn’t just about which model benchmarks better, but about who defines the security perimeter when the client wields the most power in the ecosystem.