Google lets the Pentagon use Gemini for classified work and drops out of a drone swarm contest

Google confirmed that it has allowed the Pentagon to use Gemini models in classified military systems on terms of "any lawful government purpose." The same day, it emerged that on February 11 the company had withdrawn from a US Department of Defense contest for voice-control technology for autonomous drone swarms.

What the contract covers

This is not about developing a separate weapons product, but about API access to Google's commercial models inside closed Pentagon networks. It extends existing cooperation: Gemini was previously deployed to roughly three million Defense Department personnel at the unclassified level. Access is now being extended to classified environments, which are isolated networks used for mission planning, intelligence analysis, weapons targeting and target decision-making.

The wording of the deal gives the department a very broad range of permitted use. The contract includes caveats that mass domestic surveillance and autonomous weapons without human involvement are undesirable, but they read more like a declaration than a hard technical ban. Moreover, the Pentagon can request changes to the model's safety settings and filters, while Google does not get the right to block the department's lawful operational decisions, either in real time or after the fact.

• Gemini use is permitted for "any lawful government purpose"
• Access is granted specifically in classified, isolated networks
• Google provides API access to commercial models, not a separate military model
• The contract contains restrictions in spirit, but it is unclear how they would be verified in practice

Why the dispute isn't fading

The deal was confirmed on April 28, 2026, a day after an open letter from Google employees to Sundar Pichai. More than 580 people signed it, including DeepMind researchers, directors and vice presidents. Their core argument is simple: if the models run in air-gapped networks, the company effectively cannot see which requests are sent, which responses are generated, or how the military then uses those responses inside classified systems.

"The only way to guarantee that

Google is not associated with such harm is to refuse any classified workloads."

For that reason, critics see the contract's formal caveats as weak. If a vendor cannot observe how a model is used in a closed environment, a ban on mass surveillance or on autonomous weapons without a human becomes closer to a promise on paper. For part of the Google team, the difference between "we don't build weapons ourselves" and "we provide models for classified military tasks" looks more legal than substantive. That is the core of the internal conflict.

Why Google withdrew

It also emerged that Google had advanced in a Pentagon contest with a $100 million prize pool, which called for technology to control autonomous drone swarms by voice commands. According to Bloomberg, the company notified the government on February 11, 2026 that it was no longer participating in the program, after its bid had already passed preliminary selection. Resource shortages were officially given as the reason, but an internal ethics review preceded the decision.

This is where the company's logic becomes visible. Google appears willing to sell access to general-purpose models as infrastructure, but not to directly develop a specialized system for controlling a drone swarm. In other words, the line is drawn between supplying a general-purpose tool and building a specific combat application. For legal and PR purposes this distinction matters, because it lets the company distance itself from the role of developer of a specific weapons interface.

The problem is that in practice the line blurs quickly. If the same model ends up inside classified environments where mission scenarios, intelligence and targeting are handled, the difference between "general AI" and a "military function" becomes much less obvious. That is why news of the contest withdrawal did not remove questions about the deal itself. Instead, it made the internal contradiction more visible and further fueled the dispute inside Google.

What it means

Google shows how large AI companies are normalizing military contracts: direct weapons projects can still be turned down, but access to foundation models for classified infrastructure is becoming the new norm. For the market, this signals that the debate is no longer about cooperation with defense as such, but about where the real, rather than declarative, line of control over model use lies, and whether it exists at all in closed military networks.