The Pentagon is weighing a break with AI startup Anthropic after months of talks stalled over the agency’s demand that its models be available for all lawful military uses, including weapons development and battlefield operations. Anthropic insists on retaining certain restrictions, a stance that has frustrated the defense department.

The Pentagon has urged four major AI firms—OpenAI, Google, and xAI—to lift most usage limits on classified networks. Anthropic’s refusal centers on hard caps around fully autonomous weapons and mass domestic surveillance, issues that the agency says are unrelated to current operations but critical to its strategic objectives.

An Anthropic spokesperson said the company had not discussed deploying its model Claude for specific military operations, noting that conversations have focused on policy questions rather than battlefield use. Claude, however, was reportedly deployed via a partnership with data firm Palantir during the U.S. military’s operation to capture former Venezuelan President Nicolas Maduro.

If the Pentagon follows through, it could force Anthropic to either loosen its safeguards or lose a lucrative defense contract, reshaping the competitive landscape for AI vendors in the defense sector. The dispute also signals a broader push by the U.S. military to secure unrestricted access to