Twenty-eight contracts between the Pentagon and artificial intelligence companies, according to FOIA documents

Over four hundred pages of US Department of Defense contracts have been obtained by The Intercept through litigation based on the Freedom of Information Act. The documents concern twenty-eight contracts and show that OpenAI, Anthropic, Google and xAI signed agreements in July 2025 worth up to two hundred million dollars each. The news was picked up on September 9, 2026 by specialized digests, including AI Weekly.

The news currently stems from a single source (The Intercept’s investigation, picked up by the aforementioned digest); no independent confirmation is available in the material gathered. This clarification is necessary: the content of the documents has not been verified by this newspaper, nor does it appear to have been commented on by the administrations and companies involved in the sources available.

What the agreements provide for

According to the reconstruction, the deals go beyond the provision of a service. They call for bidirectional data exchange, comparative evaluations of frontier models, engineers embedded with the military and joint exercises. This is a model of integration, not procurement: company personnel inside military facilities, and data flowing in both directions.

The most significant passage in the documents concerns operational use. According to the report, US Central Command used Anthropic technology for target identification in strikes in Iran — that is, in the ongoing conflict covered today in this newspaper’s lead story. The link between the 2025 framework contracts and a 2026 military operation is what sets this documentation apart from a mere list of IT supplies.

The model rolling out from cybersecurity

In the same week, the other end of the chain moved into public view. On September 3, OpenAI announced the start of the rollout of GPT-6 Astra, distributed in phases: the first to gain access are companies admitted to an application-based cybersecurity program called Daybreak. The announcement was reported by CNBC.

Chief executive Sam Altman described the model, in an interview with the same network, as a “new level of capability” (transl. from English). According to pricing lists tracked by industry trackers, the model is distributed under the designation gpt-6-astra with a context window exceeding one million tokens, as indicated by LLM Stats.

Caution is warranted here, and the source itself notes as much: assessments of the model’s performance and cyber-related risks are currently company statements, not independent verifications. A company that announces advanced cyber capabilities and chooses to open access first to a security program is making a distribution choice that is also an implicit statement about the product’s risk.

Why the two stories go together

Placed side by side, the two stories describe the same supply chain seen from two sides. On one hand, companies rolling out models with graduated access procedures while disclosing their risks; on the other, documents showing those same companies integrated into a state’s military decision-making chain, with embedded personnel and shared data. The two do not contradict each other: they are the structure of the market.

What we don’t know

We don’t know how much of the two hundred million dollars indicated as a ceiling for each agreement has actually been spent, nor which of the twenty-eight contracts are active today. We don’t know the responses of the four companies to the reconstructions in the documents, which do not appear in the material available. We don’t know how the technology mentioned was used in target identification, nor what role human decision-making played in the chain.

The documentary fact is the figure: twenty-eight contracts, over four hundred pages, four companies, a signing date in July 2025.