Summary: Citing a defense official, Axios reported on February 23, 2026 that the Pentagon and Elon Musk’s xAI had finalized an agreement that lets the US military run Grok inside its classified systems. The story matters because a model vendor’s safety red lines have now been converted into the openly stated price of market access. Anthropic’s Claude had stood alone as the only model authorized for the military’s most sensitive classified environments; with this deal, Grok becomes the second.
Citing a defense official, Axios reported on February 23, 2026 that the Pentagon and Elon Musk’s xAI had finalized an agreement that lets the US military run Grok inside its classified systems. Ordinary procurement news this is not. Anthropic’s Claude had stood alone as the only model authorized for the military’s most sensitive classified environments; with this deal, Grok becomes the second. And the clause that bought xAI its way in — an “all lawful use” standard — is precisely the one Anthropic had refused to put its name to.
The story matters because a model vendor’s safety red lines have now been converted into the openly stated price of market access. Sign the government’s use terms and the contract lands in your hands; refuse and you live under the threat of being cut off. What decides defense AI market share now is each lab’s tolerance for usage restrictions.
Inside the Deal: “All Lawful Use” as the Entry Fee
The original Axios report, plus February 24 follow-ups from Türkiye Today and Anadolu Agency, converges on one core fact: xAI said yes to the Pentagon’s “all lawful use” standard. Grok was already present in unclassified military systems next to Google’s Gemini and OpenAI’s ChatGPT, so the genuine news is the classified line finally being crossed.
Inside those most-sensitive environments, Claude has carried workloads spanning battlefield operations, intelligence analysis, and weapons development, which means part of the Pentagon’s AI capability is effectively Claude capability. A second model now touches that tier of defense infrastructure — a milestone more symbolic than operational for now, the ink dry but the integration work yet to begin.
What Keeps Anthropic on the Outside
Usage terms lit the fuse. Claude’s usage limits drew a threat from the Pentagon to sever the relationship with Anthropic; Anthropic answered the “all lawful purposes” demand with a refusal and dug in behind two red lines: no mass surveillance of Americans, and no fully autonomous weapons.
Per the February 24 reporting, Defense Secretary Pete Hegseth was set to sit down with Anthropic CEO Dario Amodei on Tuesday, ultimatum in hand. The heavier lever is the “supply chain risk” designation — should that label land on Anthropic, defense contractors could be compelled to cut all ties with the company, fallout reaching far beyond the Pentagon’s own contracts.
Replacing Claude: Harder Than It Sounds
Even the officials concede that the technical swap is no small job. Claude runs deep inside Pentagon workflows, Palantir collaborations included, and the reporting leaves two questions hanging: whether Grok can take over completely, and how long a transition would run. The realistic near-term shape is a two-vendor setup — Claude holds its place inside existing integrations while Grok seeps in through new workloads.
The lesson lands bluntly for product teams: government-grade AI procurement is not a one-and-done model pick but a long commitment tied to how deep the integration goes. Pulling out a model already embedded in the workflow costs roughly what rewriting that workflow would.
Google and OpenAI: Still Waiting for Their Seats
Assembled from Axios and the follow-ups, the fuller picture: Google and OpenAI are both actively chasing classified seats of their own. The New York Times had Google close to a classified Gemini deal and OpenAI further back; a defense official pushed back on that, insisting talks with both companies remain live. The common denominator is the gate itself — no company enters classified systems without first accepting the “all lawful purposes” requirement.
It is the government-side edition of the terms-tightening trend we flagged in our opening outlook for 2026 — commercial API terms and government use terms hardening in lockstep, with model vendors squeezed from both ends.
Signals for Vendors and Procurement Teams
Three points stand out. First, safety red lines now carry an explicit price tag — Anthropic put two usage limits on the table and accepted the risk of losing its exclusive classified seat, while xAI offered blanket acceptance in exchange for first-mover access; both paid, just in different currencies. Second, defense AI procurement is shifting from single-vendor lock-in to competitive multi-vendor sourcing, which hands contractors and systems integrators a fresh architecture problem: abstraction layers and swapability are turning into real bid requirements. Third, political risk has become a first-class variable in the AI supply chain — expect scrutiny of Musk’s overlapping business interests to continue, and fold policy volatility into any vendor-risk assessment that includes xAI.
For working developers, the signal is a simpler one: as capability gaps between frontier models narrow, procurement weight slides toward terms, trust, and integration depth. Picking a model vendor increasingly doubles as a policy statement.