The Pentagon is running side-by-side tests of competing artificial intelligence models with 25 of the department’s designated “power users” to determine which system could replace Anthropic’s Claude across U.S. military operations, according to a Bloomberg report published Thursday morning.
The trials began in early March, days after Defense Secretary Pete Hegseth designated Anthropic a “supply-chain risk” to national security and ordered the federal government to stop doing business with the company over its refusal to permit the use of Claude for fully autonomous weapons systems and mass domestic surveillance.
Anthropic has since filed challenges in federal courts in San Francisco and Washington seeking to overturn the designation, arguing that the blacklist could cost the company billions of dollars in lost government and commercial business if allowed to stand.
Cameron Stanley, the Pentagon’s chief digital and AI officer, told Bloomberg earlier this year that the Defense Department would require more than a month to begin transitioning away from Anthropic products already embedded inside U.S. military operations tied to the Iran conflict.
More than two months later, that transition has evolved into an active competitive evaluation process. The 25 military “power users” participating in the testing are drawn from combatant commands and intelligence operations that rely heavily on large language models for battlefield analysis, planning and operational decision support.
The companies positioned to absorb the lost Anthropic business are already emerging. On May 1, the Department of Defense announced agreements with seven major AI firms to deploy systems across classified Pentagon networks: SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, and Amazon Web Services. Anthropic was notably absent from the group.
According to the Pentagon, the participating companies share “the conviction that American leadership in AI is indispensable to national security.”
At the center of the dispute are Anthropic’s restrictions on how Claude can be used by military and government agencies. The company has maintained that it will not permit its models to be used for fully autonomous lethal targeting or mass surveillance of American citizens.
Senior Pentagon officials, including Hegseth, have argued those restrictions are incompatible with the operational demands of modern warfare.
Hegseth wrote earlier this year on X that “effective immediately, no contractor, supplier, or partner” doing business with the Pentagon could engage in commercial activity with Anthropic — an unusually sweeping designation that extended beyond direct government contracts into broader vendor relationships.
Anthropic chief executive Dario Amodei later met with White House Chief of Staff Susie Wiles and other administration officials on April 17 in what was widely viewed as an effort to ease tensions with the administration.
Following the meeting, President Donald Trump told CNBC that a deal with Anthropic remained “possible.”
“They’re very smart, and I think they can be of great use,” Trump said.
No formal resolution has emerged, and the legal fight remains active.
The financial stakes are substantial. Roughly $200 million in federal business is reportedly in question, while the administration has established a six-month timeline for agencies to migrate away from Anthropic systems.
Anthropic has also revised parts of its internal AI safety framework in recent months, shifting from binding internal scaling commitments toward a more flexible model the company says better reflects competitive realities in the global AI race.
Critics have characterized the changes as a concession designed to remain commercially competitive, while Anthropic argues that unilateral restraint by responsible developers does not prevent rivals from advancing more aggressively.
At present, Claude remains the only large language model authorized for certain classified U.S. military systems through Anthropic’s partnership with Palantir Technologies.
Competing systems including OpenAI’s ChatGPT, Google’s Gemini and xAI’s Grok are currently available inside unclassified Pentagon environments and have reportedly agreed to modified safeguard terms under their government access arrangements.
The Pentagon’s current testing process will help determine which of those systems — or combination of systems — ultimately fills Anthropic’s classified role.
The broader implications extend well beyond a single defense contract.
The Pentagon’s position effectively establishes that AI suppliers seeking government business must accept military-defined use cases without negotiating operational restrictions on a case-by-case basis.
That precedent is now forcing every major American AI developer to decide how far it is willing to go in balancing commercial opportunity, national-security cooperation and publicly stated safety commitments.
OpenAI announced its own Pentagon partnership on the same day the administration blacklisted Anthropic, with chief executive Sam Altman describing the arrangement as including “technical safeguards” accepted by the government. The specific details of those safeguards have not been publicly disclosed.
For the AI industry, the Anthropic dispute is becoming the first major test of what happens when a frontier AI company attempts to hold a published safety line against the largest government customer in the world.
If Anthropic succeeds in court, the outcome could strengthen the ability of AI firms to negotiate operational restrictions with government agencies. If the Pentagon prevails, the message to Silicon Valley will be that access to federal contracts comes on government terms — and that companies willing to remove restrictions will gain the advantage.
The Pentagon’s 25-user evaluation group is expected to issue recommendations in the coming weeks. The company ultimately selected will inherit one of the most consequential AI contracts in the federal government, while Anthropic’s path back may depend on whether the courts decide the blacklist can stand.
JBizNews Desk
© 2026 JBizNews. All rights reserved.
