The California Deal That Changes Everything

On June 29th, Governor Gavin Newsom signed the largest state-government AI deployment in history. California struck a deal with Anthropic to give every state agency, city, and county access to Claude at half price, with free training and hands-on support from Anthropic’s own developers.

The headlines will call it a procurement contract. That’s technically true and completely misses the point.

This is something stranger and more significant. California isn’t just buying software. It’s entering a partnership where Anthropic engineers will work directly alongside state workers, embedding in workflows rather than handing over a license key and walking away. The Department of Technology is already using Claude to scan state code for cybersecurity vulnerabilities. The DMV has been running a chatbot called Poppy on Claude for months. The Department of Health Care Services uses it for Medicaid workflows. None of that was announced this week — it’s been running in production. The deal announced on June 29th is the formalization of something that was already working.

The price tells you something too. Half off enterprise rates, no stated token limits, extended to all 482 California cities and counties. That isn’t a government discount — that’s Anthropic making a calculated bet on volume and stickiness. Get Claude into the workflows of 300,000 state workers and their municipal counterparts, and you’ve built infrastructure. Infrastructure doesn’t get replaced. It just gets deeper.

What makes this genuinely historic, though, is the timing. Last month, the federal government effectively banned Claude’s most powerful models — Fable 5 and Mythos 5 — citing national security concerns, before lifting those restrictions at the end of June. The Trump administration’s Department of Commerce had cut off global access to Claude for nearly three weeks. California, in the middle of all that, signed a deal doubling down.

That contrast matters. The federal government was treating Anthropic as a potential threat to be contained. California was simultaneously treating them as a strategic partner to be locked in. That’s not a contradiction — it’s a portrait of where we actually are with AI governance. There is no unified position. There is a federal government in one place, fifty states in different places, and nobody quite knows what they’re supposed to be doing with this technology except that they can’t afford to be left out.

California invited other AI companies to negotiate similar deals back in December. So far, only Anthropic came through. That matters for its own reasons — Anthropic moved fastest, and California made clear they’d prefer a relationship with a home-state company. But it also tells you something about the current landscape. Most AI companies are still building products. Anthropic is building government.

The cybersecurity angle is where this gets genuinely interesting. The California Department of Technology is using Claude Code to scan state codebases for vulnerabilities, triage incidents, and — according to the deal — patch critical systems. This is not a chatbot for writing press releases. This is an AI being used to defend critical public infrastructure.

That sentence should give anyone pause. Not because it’s wrong — state cybersecurity is catastrophically under-resourced, and any tool that helps close the gap is probably worth deploying. But because it means Anthropic, a private company, is now embedded in the security posture of the world’s fifth-largest economy. The oversight mechanisms for that are not obvious. Nobody has written the governance framework for an AI company having that kind of operational role in government.

California’s Department of Technology will route AI access through a new SITeS portal — a centralized hub for state IT services. That’s a sensible structural move. It creates visibility into who’s using what, and it gives CDT leverage to negotiate better terms over time. But it’s also a way of making Claude the default, preferred path for agencies who don’t want to figure out AI procurement from scratch.

And here’s the thing about defaults: they tend to stick.

The comparison I keep coming back to is cloud adoption in government. When AWS started winning federal contracts in the early 2010s, the conventional wisdom was that agencies would maintain relationships with multiple providers, picking the right tool for each job. What actually happened was AWS became the default, and by the time anyone got serious about multi-cloud strategies, a decade of path dependency had made that much harder. We’re in the early 2010s of AI procurement, and California just handed Anthropic the keys to a very large building.

None of this is an argument against the deal. California’s government is genuinely under-resourced, the use cases are real, and Claude has a better safety reputation than most alternatives. But it is an argument for being clear-eyed about what this is: a major infrastructure decision being made quickly, in a regulatory vacuum, by a state that moves faster than the federal government precisely because it faces fewer constraints.

Governor Newsom said it well in the announcement: “AI should not replace the human work of government. It should help our workers move faster, solve problems more effectively, and deliver better results for Californians.” That’s the right framing. Whether it becomes true depends entirely on whether anyone is actually watching whether it becomes true.

The deal is done. The next question is whether anyone is paying attention to what it means.