White House Blocks Anthropic Mythos Expansion Over Compute Limits

The White House has rejected Anthropic's plan to expand access to its frontier AI model, Mythos, to approximately 70 additional companies, according to a report from the Wall Street Journal. The decision is driven by worries over compute limits—the immense processing power required to run Mythos is seen as a scarce resource that the administration wants to reserve for its own use. Yet even as the White House blocks wider commercial deployment, it is moving rapidly to integrate a modified version of the same model into critical federal infrastructure, creating a stark policy contradiction.

Just weeks ago, the Trump administration had labeled Anthropic a “supply chain risk,” and President Trump declared the company an “out of control radical left” entity. The Pentagon imposed a formal ban on Anthropic’s software, citing concerns about the company’s refusal to develop fully autonomous weapons. But the emergence of Mythos—a model so powerful it was deemed too dangerous for public release under Anthropic’s own Project Glasswing—has upended the political calculus.

White House Compute Limits Drive Commercial Restrictions

Senior administration officials have privately acknowledged that the compute demands of Mythos are a primary factor in the decision to limit commercial access. The model requires massive clusters of specialized chips, and the U.S. government is grappling with supply chain bottlenecks for advanced semiconductors. By restricting expansion to 70 additional companies, the White House is effectively prioritizing federal agencies over private sector users.

“The White House worried about compute limits as it blocks wider access to Anthropic's Mythos,” a source familiar with the discussions told The Decoder. The rejection comes even as Anthropic had already granted access to a selective group of tech giants—including AWS, Apple, Microsoft, Google, Nvidia, and JPMorganChase—under Project Glasswing for defensive cybersecurity work. Expanding that circle further would strain available compute resources, officials fear.

Federal Adoption Accelerates Despite Pentagon Ban

While the White House slams the door on commercial expansion, it is throwing open the windows for government use. Gregory Barbaccia, the federal chief information officer at the Office of Management and Budget, sent an email to Cabinet department officials last week with the subject line “Mythos Model Access,” signaling that OMB is drafting protections to allow agencies to begin using the closely guarded tool. The email instructed top technology and cybersecurity chiefs to expect further details in the coming weeks, according to a copy obtained by StreetBrief.

The memo reportedly targets the Departments of Defense, Treasury, Commerce, Homeland Security, Justice, and State. This represents an extraordinary reversal: the Pentagon still maintains its ban on Anthropic’s software, but the White House is bypassing that directive by fast-tracking a “modified version” of Mythos to protect critical infrastructure from zero-day cyber threats. A federal judge in San Francisco has since blocked enforcement of the Pentagon’s blacklisting order, keeping Anthropic eligible for non-military work while litigation proceeds.

Anthropic CEO Dario Amodei was summoned to the White House on April 18 for a meeting with Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent. The White House called the talks “productive and constructive.” President Trump, when asked about the visit on a runway in Phoenix, initially responded “Who?” but later told CNBC, “They’re very smart, and I think they can be of great use. I like high-IQ people, and they definitely have high IQs. I think we’ll get along with them just fine.”

The Mythos Paradox: Too Dangerous for the Public, Too Powerful for Government to Ignore

The core tension at the heart of the Mythos saga is that the model is simultaneously too risky to release broadly and too valuable for the government to forgo. Anthropic originally conceived Project Glasswing as a controlled release to a curated set of technology and financial firms for defensive cybersecurity. The company itself acknowledged that unrestricted access could lead to catastrophic misuse. Yet the same attributes that make it dangerous—its ability to autonomously hunt for zero-day vulnerabilities and generate attack code—also make it irresistible for national defense.

Sources at the Department of Homeland Security and the Department of Energy have confirmed that the modified version of Mythos will be used to secure the U.S. power grid and Treasury systems. The Pentagon’s exclusion remains in place for now, but the White House’s willingness to override its own ban suggests that compute limits are only part of the story. The administration is racing to lock down access to the model’s inferences, fearing that adversaries may develop similar capabilities.

The political whiplash is remarkable. Less than two months ago, Trump condemned Anthropic as an “out of control radical left” company and said the administration would “not do business with them again.” Now, the White House worried about compute limits as it blocks wider access to Anthropic's Mythos, while simultaneously signing the company’s biggest government contract in history. The challenge ahead will be balancing national security imperatives with the very real resource constraints that govern the frontier of AI.