Last updated: September 9, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
AI giants work hand-in-hand with the Pentagon, and new contracts prove it in writing. More than 400 pages of Defense Department paperwork, obtained by The Intercept through a Freedom of Information Act lawsuit and published September 8, 2026, show OpenAI, Anthropic, Google, and xAI embedding engineers with troops and running war-game simulations. One document even shows the Pentagon asking OpenAI for models with "minimal refusal rates." This isn't a rumor. It's the contract language itself.
Short answer: OpenAI, Anthropic, Google, and xAI each signed Pentagon contracts worth up to $200 million in July 2025. They trade model access and embedded engineers for military threat briefings and war-game access. In my testing of what's public, the deals grew to eight companies by May 2026. Anthropic was frozen out in February 2026 over weapons restrictions it refused to drop.
I spend most weeks reading vendor pricing pages, not defense contracts. But this story sits right at the edge of both worlds. If you use ChatGPT, Claude, Gemini, or Grok for work, the company behind that tool is very likely a Pentagon contractor now. The terms of that deal shape what the model will and won't do, and how long the vendor stays in business. Here's what the documents say, how the four original deals split apart, and how to check your own AI vendor's government ties in about ten minutes.
What you'll need
Nothing technical. You need the name of the AI vendor your household or business pays for, ten minutes, and a browser. I'll check a vendor's defense ties the same way I check a pricing page before I recommend a tool: by going to primary sources, not a press release. If you manage AI purchasing for a team, pull your vendor list first. Then run each one through the same checklist.
Step-by-step: what the Pentagon contracts actually say, and how to check your own vendor
1. Know the baseline deal: four companies, up to $200 million each, July 2025
In July 2025, the Pentagon's Chief Digital and AI Office signed deals worth up to $200 million each with OpenAI, Anthropic, Google, and xAI. The goal: AI prototypes to "improve military advantage" and "enhance military decision making," per NBC News’ report at the time. That was the headline everyone saw. The new documents add the texture underneath it.
2. Read what The Intercept's FOIA documents actually show
The Intercept obtained 400-plus pages with help from Legal Advocates for Safe Science and Technology. The documents describe a two-way trade, not a simple vendor sale. Companies gave the Pentagon model access, briefings on "frontier AI" trends, war-game training with troops, and engineers embedded inside military units. In return, per The Intercept’s September 8, 2026 report, the Pentagon gave companies threat assessments and "risk forecasting" on their own products. That's a closer, more mutual deal than "the military bought some software."
3. Understand the "minimal refusal rates" detail, and OpenAI's pushback
One document, dated February 6, 2026 and amended February 27, 2026, describes "OpenAI Mission Models" built for "national security problem sets." Its own wording calls for "minimal refusal rates." In plain terms: a version of the model less likely to decline military requests than the public one, per The Intercept’s companion report on the same date. OpenAI told The Intercept it "has never agreed to contract language requiring 'minimal refusal rates'" and called the line draft language it rejected. Both things can be true at once. A document existed with that phrase. OpenAI says it never signed off on it as final. That gap, between a leaked draft and a signed deal, is worth watching for.
4. Track how the four original vendors split apart
The four companies that signed near-identical $200 million deals in July 2025 no longer look alike a year later. Anthropic refused to drop language blocking its models from mass domestic surveillance and fully autonomous weapons that fire without a human involved. On February 27, 2026, President Trump ordered federal agencies to stop using Anthropic's tools. Hours later, OpenAI announced its own expanded Pentagon deal, this one with an explicit ban on domestic mass surveillance and a rule requiring a human in the loop for any use of force, per NPR’s coverage that same day. Anthropic is still fighting the exclusion in court. I covered Judge Rita Lin’s skepticism of the government’s case in a separate piece, and that lawsuit is still open as of this update.
5. Check whether the circle expanded past the original four
By May 1, 2026, the Pentagon had moved well past the original four labs. It signed deals letting eight companies — OpenAI, Google, Microsoft, Amazon Web Services, Nvidia, SpaceX, Oracle, and Reflection AI — run models on classified IL6 and IL7 networks, the tiers used for secret and highly restricted data, per TechCrunch’s report on the deal, also dated May 1, 2026. Anthropic wasn't on that list. If your vendor is one of the eight, its classified-network tie to the Pentagon is now bigger than its original 2025 prototype deal, not smaller.
6. Search your own vendor's name plus "Department of Defense" before assuming anything
The fastest check: search "[vendor name] Pentagon contract" or "[vendor name] Department of War." Read the vendor's own statement next to independent reporting, the way I did for every claim in this piece. A defense tie isn't automatically bad for you as a customer. But it does mean a second, very different customer, with very different goals, is shaping how the model behaves.
Pentagon AI contracts at a glance
| Company | Original deal (July 2025) | Status as of September 2026 |
|---|---|---|
| OpenAI | Up to $200M prototype contract | Active; expanded Feb 27, 2026 deal with safeguards; one of 8 firms on classified IL6/IL7 networks (May 2026) |
| Up to $200M prototype contract | Active; one of 8 firms on classified IL6/IL7 networks (May 2026) | |
| xAI | Up to $200M prototype contract | Active as of July 2025 deal; not named in the May 2026 classified-network expansion |
| Anthropic | Up to $200M prototype contract | Federal use ordered stopped Feb 27, 2026; excluded from May 2026 classified deals; contesting exclusion in court |
| Microsoft, AWS, Nvidia, SpaceX, Oracle, Reflection AI | Not part of the original July 2025 group | Added to the classified IL6/IL7 network deal, May 1, 2026 |
Example prompts you can copy
- Check a vendor's government ties: "Search for [AI vendor name] and 'Department of Defense' or 'Department of War' contract. Summarize what the contract covers, its value, and the date, citing the source."
- Compare two vendors' restrictions: "Compare what [Vendor A] and [Vendor B] have publicly said about restricting their AI models from autonomous weapons or mass surveillance use. Cite each company's own statement."
- Track a live legal dispute: "Search for the current status of [company]'s lawsuit against [agency]. What was the most recent court ruling, and when is the next hearing?"
- Stress-test a claim before you repeat it: "Here's a claim I read: [paste claim]. Find the original source, its publish date, and whether any party involved has disputed it."
Common mistakes to avoid
The mistake I see most is treating "the Pentagon signed a contract with X" as proof that X is endorsed, or proof that X is compromised. In my testing, neither reading holds up once you read the actual document. A second mistake: missing that these four companies no longer look alike. Anthropic and OpenAI signed very different follow-up deals within the same 24-hour window in February 2026 (per NPR's report above), and lumping them together erases the distinction that mattered most. Third, don't treat a leaked draft clause, like the "minimal refusal rates" language OpenAI disputes, as an automatically-signed final term. Check whether the company responded on the record before repeating it as fact. Fourth, don't assume this only matters if you work in defense. If the model you use every day is also being tuned for military tasks under a separate contract, that shapes how the vendor spends engineering time and honors safety promises across its whole product line.
Tools that make this easier
For the deeper legal detail on the Anthropic side, my breakdown of Judge Lin’s ruling on the Anthropic ban covers both active court cases and what they mean for Claude users. For the wider pattern of governments betting big on a handful of AI labs, see why governments are making a dangerous bet on the AI boom. If xAI's defense and infrastructure ties are what you're after, xAI, SpaceX, and the race for AI buildout covers Musk's overlapping companies in more depth. For how OpenAI and Anthropic each frame their own safety promises, my comparison of AI responsibility at OpenAI and Anthropic is the place to start. Anthropic’s $30 trillion revenue projection shows how much rides on these companies' wider business, not just government deals. Once you've checked your own vendor's ties, my AI tool ratings hub and free AI tools roundup are where to compare alternatives. AI is getting way too expensive covers the pricing side of the same vendor-concentration risk.
My take
None of this means you should drop an AI tool because its maker also has a Pentagon contract. Nearly every major lab now does, in one form or another, and that fact alone tells you less than the specific terms attached. What I'd actually watch is the gap between a company's public safety language and what its government contracts require. That gap is exactly what split Anthropic from OpenAI in a single February afternoon. Read the primary document when one exists, not just the headline about it. Recheck any vendor you depend on for work at least twice a year. This list changed twice in the last twelve months alone.
Frequently Asked Questions
Did AI companies really sign contracts with the Pentagon?
Yes. OpenAI, Anthropic, Google, and xAI each signed contracts worth up to $200 million with the Department of Defense in July 2025, and the Pentagon added Microsoft, Amazon Web Services, Nvidia, SpaceX, Oracle, and Reflection AI to a separate classified-network deal in May 2026.
Is Anthropic still banned from working with the Pentagon?
As of this update, yes — federal agencies were ordered to stop using Anthropic's technology on February 27, 2026, and Anthropic was left off the May 2026 classified-network deal. Anthropic is contesting the exclusion in court, and a federal judge has questioned whether the government has justified it.
Did OpenAI actually agree to build AI with "minimal refusal rates" for the military?
A February 2026 contract document used that phrase for a national-security-focused model, but OpenAI told The Intercept it never agreed to that specific language and called it draft material it rejected. The document exists; whether it became final signed policy is disputed.
Does my Claude, ChatGPT, or Gemini subscription change because of this?
No. These are separate federal-government contracts distinct from the consumer and business products most people pay for. None of the pricing or features on the public plans changed because of these deals.
How long does it take to check whether my own AI vendor has Pentagon ties?
About ten minutes per vendor: search the company name plus "Department of Defense" or "Department of War," read one or two independent reports alongside the company's own statement, and note the date — these relationships have changed multiple times within a single year.