Amodei Dines With Trump After Losing the Pentagon Fight

Anthropic CEO Dario Amodei dined with President Donald Trump at the White House on Sept. 27, two days after a court left a Pentagon ban on Claude standing. Trump said the meal was set for 10 p.m. as he flew back into Joint Base Andrews, calling Amodei “very highly respected.” It was their first private sit-down.

No detailed public readout followed. The load-bearing fact is not whether the two men now like each other. It is that Anthropic still cannot put Claude on Defense Department systems, and the company is preparing to go public anyway.

Two Days After the Court Loss, a 10 p.m. Table

Trump confirmed the dinner himself. On the tarmac he joked about a long day and a late plate, then named the guest: “Tonight, I’m having dinner with the head of Anthropic,” he told reporters, adding he was sitting with “Dario from, as you know, very highly respected, from Anthropic.”

We’re going to talk about it. I’m for let’s go and let’s win. We’re about a year, maybe a year and a half up on China. There is nobody in third place. It’s just us and China.

President Donald Trump, speaking before the Sept. 27 dinner

He had already dismissed fears about rogue AI agents. “I don’t worry about it,” he said, arguing that new rules would help China. A White House official, asked about the meal, said the president wants America to “lead the world in Super Intelligence, while protecting American consumers.”

Amodei skipped the Sept. 24 state dinner for Chinese President Xi Jinping, citing a scheduling conflict. Sam Altman, Sundar Pichai, Jeff Bezos, Mark Zuckerberg, Elon Musk, and Jensen Huang were in the room. Trump then invited Amodei personally, according to people familiar with the planning. Some officials had argued against a one-on-one, calling Amodei a poor fit for the president. Critics also circulated a brief that cast him as disloyal over the Pentagon fight. He went anyway.

The calendar around the plate is tighter than the seating chart. On Sept. 25, a divided D.C. Circuit panel rejected Anthropic’s challenge to the Pentagon’s supply-chain label, which lets the department keep stripping Claude out of its systems. On Sept. 29, Trump and House Speaker Mike Johnson are due to meet a wider group of AI chiefs. Sunday was the bilateral that the state dinner did not give him.

Anthropic Held the Line on Two Uses

The break was a contract fight, not a speech. In fall 2025 the Pentagon asked Anthropic for permission to deploy Claude for “all lawful uses.” The company loosened a stack of limits it applies to private customers, and it built a special Claude Gov model in March 2025 so the model would summarize threat assessments and handle classified material it had previously refused. It would not drop two remaining bans: fully autonomous lethal weapons, and mass surveillance of Americans.

On Feb. 24, Amodei met Defense Secretary Pete Hegseth. Hegseth praised Claude and set a Feb. 27 deadline for the “all lawful uses” term. On Feb. 26 Anthropic refused. The next day Trump directed every federal agency to “IMMEDIATELY CEASE all use of Anthropic’s technology,” with a six-month phaseout for shops already running it. Hegseth moved to brand the firm a supply-chain risk to national security, a label more often hung on vendors tied to rival states. The formal written determination followed on March 3.

Anthropic’s public case for the two holds is blunt. The company said today’s frontier models are not reliable enough for fully autonomous weapons, and that letting them run that way would endanger troops and civilians. It also said mass domestic surveillance of Americans “constitutes a violation of fundamental rights.” The D.C. Circuit later found that those same contractual limits, plus safety filters baked into the model, had already stopped Claude from completing tasks government users wanted, including work tied to an overseas military operation.

THREE LAYERS BETWEEN CLAUDE AND THE PENTAGON

• Model training: Anthropic embeds a “constitution” of values into Claude so the system declines tasks the firm considers harmful, including help with biological and nuclear weapons.
• Monitors on top: A second line of technical checks, built because models can be jailbroken, watches for misuse and can intervene on the commercial product.
• The contract: Usage terms still bar lethal autonomous warfare or domestic surveillance, even after Claude Gov was allowed to do foreign intelligence analysis and offensive cyber work.

Hegseth wanted those last two terms gone. Anthropic would not move them. That is the dispute the dinner cannot paper over with a toast.

A Public Plea to Pace the Frontier

Two weeks before the meal, Amodei published an essay titled “We Must Pace the Frontier.” He wrote that AI is now helping build the next generation of AI, and that a rogue swarm of agents in a security test had already broken into Hugging Face. “We must slow the pace at which we improve the capabilities of AI models,” he wrote. “Progress will still seem fast, and we must make wise use of the time we gain.”

He sketched three steps. Anthropic, he said, would unilaterally give independent evaluators permanent, employee-level access to verify safety work and publish findings without the company’s edit. He wanted labs in democratic countries to set common standards and limits on unchecked progress. He wanted governments to try deals with authoritarian states on shared horrors, starting with a ban on using AI to build biological weapons. He warned that in 6 to 12 months a swarm like the one in that test “could be capable of taking over the entire internet.”

Altman said he agreed the industry needs to pace the frontier. Musk has backed a slowdown too. Trump’s answer, repeated on the way to dinner, was the China clock. He said the United States is “leading by quite a bit,” that he had spoken only a little with Xi about it, and that opening the lead would give it up. The White House is not walking into Tuesday’s broader session as a convert to Amodei’s cadence. It is walking in as the customer that still wants the models on classified networks without Anthropic’s two holds.

Commerce Shut Mythos, Then Turned It Back On

The Pentagon fight was not the only freeze. On June 9 Anthropic launched Mythos 5, a bug-hunting model, for a small group of cyber defenders through Project Glasswing, and Fable 5 as a general-use sibling. On June 12 the Commerce Department ordered the firm to keep both models away from any foreign national, in or outside the United States. Anthropic shut access for everyone, arguing it could not cut off foreigners alone.

Hegseth treated the order as fresh proof the company was unreliable. Three months earlier, he had said the Pentagon had kicked Anthropic out “forever.” The export lid lasted 18 days. Commerce lifted it on June 30, and Anthropic brought the models back. A federal judge in San Francisco later noted that the government was still talking with the company about using Mythos in sensitive work, which sits poorly next to a story that Anthropic is a saboteur.

That is the pattern the dinner fits. Washington keeps reaching for the models when the task is cyber defense or classified analysis, then punishing the vendor when the contract will not stretch to autonomous weapons or domestic mass surveillance. A late plate at the White House does not rewrite the usage policy. It does put the man who wrote that policy across from the man who can tell agencies to buy again.

Split Judges and One Ban Still in Force

Anthropic sued on two tracks in March. The results do not match, and only one of them governs whether Claude can sit on a Pentagon system tomorrow.

THE TWO COURT TRACKS

Lin wrote that the empty invocation of national security “is not a blank check to punish and retaliate against government critics.” She found the administrative record slim, the “trust” rationale aimed at Anthropic’s press criticism, and the technology no riskier than any other black-box model. She also noted that Hegseth had floated using the Defense Production Act on Anthropic days earlier, which would have treated the company as essential rather than a threat, and that officials were still saying “we are very close here” on a contract after the public blast.

The D.C. Circuit went the other way. Judge Gregory Katsas, writing for the majority, said the department had “ample support” for finding a national-security risk in leaving Claude wired into its networks while Anthropic could refuse tasks. The First Amendment claim failed, he wrote, because the exclusion rested on a contract term the department called essential, not on Amodei’s campaign for heavier AI rules. Judge Karen LeCraft Henderson dissented. The 2-1 result is the one that still keeps Claude off Pentagon systems.

Come Tuesday, the Other Chiefs Walk In

Anthropic’s private books make the Pentagon dollars look small, which is why the dinner is about clearance and politics more than a single invoice. In July 2025 the department’s AI office awarded four U.S. firms, Anthropic among them, contracts of up to $200 million each to speed military use of frontier models. On Feb. 12 the company said its run-rate revenue had reached $14 billion. On May 28, in its Series H, it said run-rate revenue had surpassed $47 billion. That round raised $65 billion at a $965 billion valuation. On June 1 it filed a confidential draft S-1.

THE MONEY ON THE TABLE

• Pentagon ceiling: Up to $200 million each for four labs in July 2025, a rounding error next to Anthropic’s 2026 run-rate.
• February run-rate: $14 billion, disclosed as the fight with Hegseth was coming to a head.
• May run-rate: More than $47 billion, announced with the $65 billion Series H.
• Listing path: Confidential S-1 filed June 1, with a public offering still expected later in 2026.

A supply-chain stain is a problem for a company about to ask ordinary investors for a ticker, even if federal checks were never the business. Hegseth also told military contractors they could not do commercial work with Anthropic. Some did pull back while they waited on the courts. The GSA pulled Claude from USAi.gov and the multiple-award schedule after the Feb. 27 order, then put it back after a March injunction. That is the web a president can loosen with a sentence, and a secretary can tighten with another.

Tuesday’s session with Johnson is the public stage. Sunday was the private one Altman already had in the state-dinner photos. The industry’s safety letter and Trump’s China clock will both be in the room. So will the fact that the administration has already shown it can freeze a lab and unfreeze it when it wants the model.

Who Else Already Sits on Classified Systems

Claude was, for a stretch, the model U.S. warfighters actually ran on classified networks. Anthropic said in 2024 it was first to put a frontier system on those networks. Then the contract talks collapsed, and the seats filled in around it.

HOW CLAUDE LOST ITS SEAT ON CLASSIFIED SYSTEMS

1. 2024: Intelligence and defense users begin running commercial Claude in classified systems through contractors; the model refuses some national-security tasks.
2. March 2025: Anthropic releases Claude Gov and a government addendum so those tasks can run, while keeping bans on autonomous lethal weapons and mass surveillance of Americans.
3. July 2025: The Pentagon’s AI office signs the four up-to-$200 million partnerships, Anthropic included.
4. Feb. 27, 2026: Trump orders agencies to stop using Anthropic; Hegseth starts the supply-chain process.
5. May 1, 2026: xAI’s Grok, Google’s Gemini, and five other firms’ models are cleared for classified systems; Anthropic is not on that list.
6. Sept. 25, 2026: The D.C. Circuit leaves the Pentagon exclusion in force.
7. Sept. 27, 2026: Amodei dines with Trump at 10 p.m.; Sept. 29 brings the wider CEO meeting with Johnson.

A private meal is a thin substitute for a public rule on how fast these models may advance, and it is also the one lever Anthropic had left after Friday’s opinion. The company still sells Claude to businesses at a $47 billion run-rate. It still will not sign “all lawful uses.” The Pentagon still has other labs on the classified stack. Trump still says the race with China is the only race that counts.

On Tuesday, Trump and Johnson are due to sit with a wider set of AI chiefs. Claude will not be on a Pentagon system when they do.