Anthropic and OpenAI agree on almost nothing when it comes to how AI should be regulated in the United States, and July 2026 made that split impossible to ignore. Anthropic became the first AI company to formally endorse California's S.B. 1053, then doubled down on a strategy its own head of U.S. state and local government relations describes in explicitly competitive terms: encourage each new state bill to raise the bar higher than the last one. OpenAI, meanwhile, is telling anyone who'll listen — including The Hill — that it wants a single, streamlined national framework, built by cherry-picking a handful of major state bills and treating them as the de facto national standard. Two of the most influential labs shaping the next decade of AI policy are now pulling in opposite directions, and the enterprises deploying their models don't get to sit this one out.
The split, spelled out
Start with what actually happened. Anthropic endorsed California's S.B. 1053, arguing the bill mostly formalizes AI safety testing practices the company — and others in the industry — already perform internally. That's a notable move: being first to back a state bill is a statement of intent, not just a policy footnote. Anthropic also endorsed New York's RAISE Act and Illinois's S.B. 315, and on those two, OpenAI is right there with them. So this isn't a case of the two labs disagreeing on every individual bill. It's something more structural: Anthropic wants the pace and stringency of state AI law to keep climbing, bill after bill, while OpenAI wants the current crop of state laws to become the ceiling, not the floor, of what companies have to comply with.
Illinois is a useful data point for why this matters beyond the abstract. The state recently mandated annual AI audits — a real, recurring compliance obligation, not a symbolic gesture. That's the kind of requirement that shows up in a legal team's calendar every year from now on, and it's exactly the sort of provision Anthropic wants to see replicated and intensified elsewhere, and that OpenAI wants contained to a manageable set of jurisdictions rather than spreading unchecked.
Anthropic's bet: make every state bill tougher than the last
Anthropic's public position, laid out by its head of U.S. state and local government relations, is unusually candid for a company its size: it wants successive state AI bills to add stronger duties over time. The word used internally is "one-upmanship" — a deliberate strategy of encouraging states to compete with each other on how tough their AI guardrails are, rather than converge around one common, static set of rules. That's a bet that regulatory competition between states produces better outcomes than regulatory harmonization, and it's a bet that only makes sense if you're confident your own practices already clear whatever bar comes next.
That confidence shows up directly in how Anthropic framed its S.B. 1053 endorsement: the bill formalizes testing practices it already does. In other words, Anthropic isn't volunteering to be regulated up to some new standard — it's asking regulators to codify what it already treats as baseline responsible behavior, then betting that competitors who haven't invested as heavily in safety testing infrastructure will have a harder time keeping up as the bar rises in state after state. Whether or not that's the primary motivation, it's a coherent one, and it explains why Anthropic isn't just tolerating a fragmented state-by-state landscape — it's actively cultivating it.
The company has also drawn a specific line on federal preemption: it does not believe Congress should override state AI law unless it replaces it with a federal law that is at least as strong as what's already being proposed at the state level. That's a meaningful qualifier. It's not opposition to federal legislation in principle — it's opposition to federal legislation used as a ceiling-lowering mechanism, where a weaker national standard preempts stronger state rules and leaves companies with less obligation than they'd have had otherwise. Anthropic's stated reasoning is blunt: a government response to AI risk "can't wait for action in Washington." Given how long comprehensive federal tech legislation has historically taken to pass — and how contested any federal AI bill would be — that's less a philosophical stance and more a statement about timelines.
OpenAI's bet: pick the winners and call it done
OpenAI's chief of global affairs laid out the opposite theory of the case to The Hill: rather than let all fifty states run their own experiments, OpenAI wants to identify a handful of the more significant state AI bills already in motion and effectively promote them into a national standard by adoption and pressure, without waiting for Congress to formally legislate. OpenAI's own term for this is "reverse federalism" — instead of the federal government setting a floor that states build on top of, a small number of influential states set the standard and the rest of the country backs into it informally.
The appeal of that approach is obvious if you're running compliance for a company operating across all fifty states: fewer standards to track, more predictability, and a regulatory environment that stabilizes faster. It's also, not coincidentally, an approach that gives large, well-resourced labs more control over which state standards become the de facto national ones — since a "handful of major state bills" doesn't get selected by accident. Whether that's a cynical read or just an accurate description of how influence works in practice, the practical effect for OpenAI's own compliance obligations is the same either way: fewer moving targets, and more certainty for the model providers that have to build compliance tooling against those targets in the first place.
"Wait for Washington" is not a strategy right now
The most important thing about this divergence isn't which lab is right — it's what it tells you about the near-term future of AI regulation in the U.S. There is currently no comprehensive federal AI law. In its absence, individual states are legislating on their own timelines, and two of the most consequential companies operating in this space are actively pulling that patchwork in opposite directions rather than working to consolidate it. Anthropic wants the patchwork to keep getting denser and stricter. OpenAI wants a subset of it to harden into a de facto standard. Neither of those outcomes looks like "a single federal law arrives soon and simplifies everything for you."
That matters enormously for any organization treating "we'll figure out compliance once Congress acts" as an actual plan. Congress has shown no near-term sign of producing comprehensive AI legislation, and even if it did, Anthropic's own stated position is that it would only support preemption if the federal standard matched or exceeded the toughest state rules already on the books — which would still leave you dealing with stringent, evolving obligations, just with one signature instead of fifty. Waiting for federal clarity isn't a hedge against complexity; it's a bet that complexity will resolve itself in your favor, on a timeline nobody involved is promising.
What this actually costs a multi-state enterprise
If you operate in more than a handful of states — which, for anything with a website, is most companies — you're now implicitly subject to a growing set of state-specific AI obligations that don't map cleanly onto each other. Illinois wants annual audits. California is moving on testing-practice formalization via S.B. 1053. New York's RAISE Act and Illinois's S.B. 315 both have industry backing from labs that otherwise disagree on regulatory philosophy, which is itself a signal that these two bills in particular are likely to stick and to be used as templates elsewhere.
The practical burden isn't just "read more bills." It's that different states are legislating different things — audit cadence, testing disclosure, safety documentation, incident reporting — and a company's AI governance program has to satisfy the union of all of them, not the lowest common denominator. Worse, the two largest model providers you might be building on top of have opposing incentives about whether that union keeps growing (Anthropic) or gets frozen around a manageable core (OpenAI). Your vendor selection is no longer just a technical or cost decision — it's tacitly a bet on which regulatory future you're more exposed to, because your vendors are actively lobbying to bring that future about.
Practical takeaways
- Stop treating "no federal AI law yet" as a reason to delay governance work — state obligations are accruing now, with real deadlines like Illinois's audit cadence, regardless of what Congress does.
- Build a live tracker of AI-specific state legislation in every state where you operate or have users, not just your state of incorporation — S.B. 1053, the RAISE Act, and S.B. 315 are three names to start with, but the list will grow.
- Assume the compliance bar rises over time rather than stabilizes, given that one of your model providers is explicitly strategizing for exactly that outcome.
- Ask your AI vendors directly which regulatory posture they're lobbying for — a lab pushing for stricter, more numerous state rules has different downstream compliance implications for you than one pushing to consolidate around fewer standards.
- Don't assume a bill one lab endorses represents industry consensus. RAISE Act and S.B. 315 have support from both labs; S.B. 1053 so far has Anthropic alone. That gap is informative about where friction is likely.
- Loop legal and compliance into AI vendor selection meetings, not just security and engineering — vendor regulatory strategy is now a material factor in projecting your own compliance workload.
- Revisit your AI governance documentation cadence. If audit and testing-disclosure requirements are expanding state by state, an annual review cycle built for a pre-2026 regulatory environment is probably already out of date.
There's a second-order effect worth naming too: when the two most prominent labs in a market both lobby heavily, but for opposite outcomes, they end up legitimizing state-level AI legislation as a category regardless of which side wins any individual fight. Every time Anthropic or OpenAI publicly endorses a state bill, it makes that bill harder for other companies to dismiss as fringe overreach, and it makes the next state legislature more confident that serious industry engagement — not just industry resistance — is the norm for this kind of lawmaking. That dynamic alone should push the "we'll wait and see" crowd toward "we should have a position and a compliance plan," because the volume of state activity here isn't slowing down under either company's preferred future.
None of this resolves cleanly, and it isn't supposed to. Anthropic and OpenAI aren't disagreeing about a technical detail — they're placing different bets about what kind of regulatory environment produces safer, more accountable AI at scale, and each bet happens to align conveniently with each company's own competitive position. That's not a criticism unique to either lab; it's how regulatory lobbying works everywhere. What it means for you is that the "wait and see" compliance posture that might have made sense a year ago is no longer defensible. The patchwork is the terrain now, not a temporary inconvenience on the way to something simpler.