The reversal nobody saw coming.

The Trump White House spent the back half of 2025 making Anthropic into the bad guy. The Pentagon labeled the company a supply-chain risk. Federal agencies were told to stop onboarding Claude. Anthropic sued the administration in March, and the case is still alive.

Now the same White House is quietly drafting a way out.

Axios first reported Tuesday that the administration is preparing executive guidance that would let federal agencies sidestep the supply-chain designation and use Anthropic's most powerful model, Mythos. Reuters, Storyboard18, and Nextgov/FCW confirmed the reporting through the day. Nextgov flagged the cyber-focused Mythos variant specifically as the model agencies want access to.

Translation: the policy outlasted the politics. Mythos is too good to ignore, and the people inside the federal government know it.

What actually changed.

The supply-chain designation came down in February. Anthropic refused to remove the constitutional guardrails the Pentagon wanted stripped from Claude on classified networks. OpenAI, xAI, and now Google all said yes to the same request, with various levels of public red lines. Anthropic was the only one of the four major American AI labs that walked.

The administration framed Anthropic's stance as ideological. The Pentagon framed it as a procurement problem. Anthropic framed it as a constitutional one and went to federal court. Two months later the company secured a $40 billion investment from Google, $5 billion from Amazon, and an October IPO target north of $350 billion.

The argument the White House lost wasn't legal. It was operational. CISA needed Mythos for cyber. Treasury wanted it for fraud detection. The intelligence community wanted it for analyst summarization. Every federal CIO with a budget kept asking the same question: why can we not use the model that benchmarks beat everything else on the market?

The Mythos exception.

The draft guidance, per Axios's reporting, would create a structured carve-out. Agencies that want to procure Mythos would file a justification, the supply-chain designation would be waived for that specific use case, and the model would clear acquisition. It is the federal-procurement version of a hall pass.

Two things to notice. First, this isn't a reversal of the Pentagon's position. The administration is not removing the supply-chain risk flag. It's building a tunnel around it. The political face-saving is intact. The contracts can flow.

Second, the guidance specifically names Mythos. Not Claude broadly. The model that earned the public security disclosures, the Project Glasswing partnership, and the largest qualitative AI study ever published is the one the administration cannot afford to keep out of federal hands.

What this does to the lawsuit.

Anthropic's federal court complaint argues the supply-chain designation was retaliation for the company refusing to remove safety guardrails. If the White House publishes guidance that lets agencies use the model anyway, the lawsuit's damages case gets harder to argue. You can't simultaneously claim you were locked out of federal contracts and accept a procurement workaround that gives you the same contracts.

Expect Anthropic to drop or settle the suit within 30 days of the guidance landing. The IPO calendar makes it inevitable. You don't go public in October with active federal litigation when the federal government just gave you the keys.

Who loses here.

The hawks inside the administration who pushed the original ban. They picked a fight with a company whose product turned out to be irreplaceable, and they lost the internal argument to the procurement officers who actually have to run things.

And the message to every other AI lab is now extremely clear. Refuse to remove your safety guardrails. Sue the federal government. Wait six months. The market will route around the policy and the policy will route around itself.

Anthropic just won the safety argument by being too useful to lose. That is a brand outcome OpenAI cannot buy with capex.

What to watch next.

The guidance text itself. Whether the carve-out is Mythos-specific or covers the full Anthropic stack. Whether xAI and OpenAI try to claim the same workaround for their own model variants. Whether the Pentagon updates its February designation memo or simply ignores it. And whether Anthropic's October IPO road show now leans into the federal contract pipeline as a revenue line, not a litigation risk.

We've been writing this story since February. The Pentagon's classified deal with Google last week was supposed to seal Anthropic's exclusion. Eight days later, the White House is opening the back door.

Sometimes the safety-first lab wins. It just takes a while.