SAN FRANCISCO — The version of Claude that the U.S. Department of Defense called too dangerous to give a government contract will now handle customer inquiries at the California DMV, assist Medicaid case workers navigating the largest state health program in the country, and scan state cybersecurity systems for vulnerabilities.
On June 29, Governor Gavin Newsom announced a deal with Anthropic giving every California state agency and local government access to Claude at a 50 percent discount, along with free workforce training and on-call developer support from Anthropic engineers. The agreement makes Claude the first AI productivity tool offered statewide through California’s centralized software purchasing portal. What the announcement did not mention, and what California’s chief information officer made clear when asked, was that the Trump administration’s designation of Anthropic as a supply-chain risk simply “didn’t come up” during negotiations.
That framing tells you more about the current state of AI policy in the United States than the deal itself does.
Earlier this year, Anthropic and the Pentagon fought over a clause that would have let the Defense Department deploy Claude for any lawful purpose. Anthropic refused. The company was willing to adapt its usage policies for government work, but it held two lines: no fully autonomous lethal weapons systems, no mass domestic surveillance of American citizens. Defense Secretary Pete Hegseth declined, signed contracts with eight companies — OpenAI, Google, Microsoft, Amazon Web Services, Nvidia, SpaceX, Reflection AI, and Oracle — and left Anthropic out entirely. The administration then escalated, designating Anthropic a supply-chain risk, a classification historically applied to companies tied to foreign adversaries, and ordering federal agencies to cease Claude use immediately. Export controls on Anthropic’s most advanced models followed, with the Commerce Department restricting international access for nearly three weeks before reversing course on July 1 after its own factual basis fell apart. Anthropic has since filed suit, and a federal judge in California has blocked the government’s effort pending that litigation.
California opened deal talks shortly after.
The terms of the agreement cover a broader footprint than most enterprise software contracts. Every state agency, city, and county in California is eligible, with access routed through the California Department of Technology’s Statewide Information Technology Shared Services portal, a new centralized hub designed to standardize AI procurement across state government. The 50 percent discount applies to Claude’s enterprise tier, bundled with Anthropic developer assistance for workflow integration. No total cost figure has been disclosed. The deal is structured as an offer to the state, not a mandate for any individual agency.
Several agencies were already using Claude before June 29. The Department of Motor Vehicles deployed it to reduce call center wait times. The California Department of Health Care Services, which runs the largest Medicaid program in the country, integrated it into case management workflows. The state’s technology department and the California Office of Emergency Services are running Claude for cybersecurity scanning, using it to identify and patch vulnerabilities in state code. A tool called Poppy — built on Claude, piloted across 67 state departments and roughly 2,800 employees — is on track for full statewide rollout this month.
Anthropic’s head of Americas, Kate Jensen, framed the deal in terms of mission alignment. “As a California company, we feel a real responsibility to our home state,” she said. “Building AI responsibly and in service of people has been our approach from the start, and that’s exactly what this partnership puts into practice.”
That formulation is doing real work. The same usage restrictions that cost Anthropic its Pentagon contract are what California’s government is now, implicitly, endorsing. No autonomous weapons deployment, no mass surveillance: those conditions apply differently when the customer is a state Medicaid agency and a motor vehicles department rather than a branch of the armed forces. The dispute over whether an AI company gets to draw those lines at all — and what happens to a company that tries — is unresolved in courts and in Congress simultaneously.
Newsom has positioned California as the counterargument to Washington’s approach to AI regulation, signing executive orders on safety and worker protection while the federal government has moved to deregulate. He framed the deal in terms of limitation as much as capability. “AI should not replace the human work of government,” he said. “It should help our workers move faster, solve problems more effectively, and deliver better results for Californians.” That construction — should not replace, should help — is not neutral territory in mid-2026, when state employee unions are watching AI procurement closely and the governor’s office is simultaneously directing agencies to prepare for AI-driven job displacement. His office described the agreement as “not intended as a political statement,” which is the kind of formulation you offer when the politics are unavoidable.
The federal government has been deciding which AI products get released at all: the Trump White House restricted initial access to OpenAI’s most powerful new model to roughly 20 pre-approved companies. The logic — who controls access to frontier AI, under what conditions, through which institutions — is where the federal and state frameworks are most directly in tension. OpenAI, which now holds contracts to run models inside federal classified networks, has been rewarded for accepting those conditions. Anthropic, which refused them, is now California’s partner instead.
What neither Newsom’s office nor Anthropic has clarified is what independent oversight applies when Claude is used in decisions with real stakes: a Medicaid eligibility assessment, a DMV record check that touches law enforcement, a cybersecurity finding that shapes a state agency’s response to an active threat. The state’s existing data privacy and cybersecurity framework governs usage, according to the announcement, without specifying what that means in practice for AI-assisted government decisions. OpenAI faces similar accountability questions as it integrates deeper into federal systems — questions neither company has answered publicly, and that no regulatory body has yet required them to answer.
California’s calculation is that the company the federal government banned is actually the safer choice for state workers and residents. Anthropic’s calculation is that the largest state in the country counts for more than losing the Pentagon’s supply chain. Whether those bets pay off depends on a federal lawsuit still working through the courts, export control negotiations that reversed once already, and a framework for AI accountability inside government that no one — in Sacramento or Washington — has yet written down.
