Dario Amodei, Anthropic's CEO, published an essay on Wednesday, June 10 arguing that the US government should be able to block or reverse the release of frontier AI models that fail independent safety tests. The company paired the essay, Policy on the AI Exponential, with two formal frameworks and a $350 million funding pledge aimed at the labor-market damage Amodei expects AI to cause.
For roughly three years Anthropic's public line was transparency: make labs disclose their safety procedures, hold off on binding law until the risks got concrete. That position is gone now. What replaced it is a request for the government to hold legal authority over Anthropic's own products, which is not a thing AI companies usually ask for.
From transparency to a release gate
The model Amodei reaches for is the FAA. Frontier models, like airplanes, should have to pass technical testing before they ship, and the government should be able to ground the ones that fail. The Advanced AI Framework spells out the mechanics, and it is more specific than these documents usually get.
It would apply only to models trained on more than 10^25 floating-point operations, built by companies pulling in over $500 million in AI revenue or spending more than $1 billion on research. In other words, a handful of labs, Anthropic among them. Qualifying models would face mandatory third-party testing across four risk categories: biological weapons, cybersecurity, loss of control, and automated AI research that could speed up the other three. Fail the test in a way that signals catastrophic risk, and the government could block deployment, with civil penalties scaled to global revenue.
Amodei points to Claude Mythos Preview, Anthropic's restricted cyber-capable model, as his evidence that these systems are now strategically consequential. The framing does a lot of work here, since the proof that strict rules are needed happens to be a model only Anthropic has.
The money, and what it's measuring
The companion Economic Policy Framework maps out three unemployment scenarios: around 5 percent, 10 percent, and an "unprecedented" level. Each tier unlocks bigger interventions, from wage insurance and retraining grants at the low end to basic income and sovereign-wealth-style equity sharing at the top, which the company admits is "novel economic territory."
Worth holding against that: the latest US unemployment rate came in at 4.3 percent. So the entire ladder starts above where things actually are right now, which makes the whole exercise a bet on a future that hasn't shown up in the data yet.
The $350 million splits into a $200 million Economic Futures Research Fund and a $150 million fellowship for early-career Americans. The Associated Press noted that scant details were available the day of the announcement, so treat the number as a commitment rather than a plan. A pledge with no published figures attached to specific programs is still mostly a press release.
Is this just a moat?
The reaction split fast. A widely shared critique from Kingy AI argued the framework risks turning safety into a moat for incumbents, since compute thresholds and accredited-evaluator regimes are easiest to absorb for the labs that already run red teams and produce compliance paperwork. Startups training their first big model do not have that infrastructure.
Amodei tried to answer the charge inside the essay itself, writing that people worry about AI because the risks are real, "not because AI CEOs have been insufficiently Panglossian." Maybe. But the timing makes skepticism easy: the essay landed one day after Claude Fable 5 shipped, a launch that already drew complaints over token costs and a 30-day data-retention mandate. Critics now have a ready example of safety rules raising the bill for users.
And there's the IPO. Anthropic is heading toward a public offering after a recent funding round, which gives every "we need a regulator with teeth" line a second possible reading.
Anthropic wants the economic agenda discussed at the G7 and an upcoming AI summit in Geneva. Whether Congress touches the testing-and-blocking regime is a separate question, and a colder one. A June executive order already lets the government vet the most advanced models for up to a month before release; Amodei calls that welcome but not enough.
