Anthropic and the Trump administration reached a deal on Friday to bring one of the company’s most powerful artificial intelligence models back online, after days of tense negotiations.
In a letter sent to Anthropic, the Commerce Department gave the San Francisco company permission to restore some clients’ access to its Mythos 5 model, which the government restricted two weeks ago because of national security concerns.
Talks were still underway to restore access to another powerful Anthropic model, Fable 5, and to provide Mythos to more organizations, said a person familiar with the talks, who spoke on the condition of anonymity because the person was not allowed to discuss them publicly.
“Anthropic has worked with the U.S. government to address risks associated with the covered models,” said the letter, which Commerce Secretary Howard Lutnick sent to Anthropic on Friday. “These efforts have yielded significant progress.”
Mr. Lutnick added that Anthropic had committed to working with the U.S. government on protocols for future releases of A.I. models.
A spokesperson for Anthropic said the company was working to restore access to the clients approved by the administration and was “pleased to see this progress and continue to work with the government.”
The letter was reported earlier by Semafor.
The deal de-escalates the feud between the Trump administration and Anthropic, one of the most influential A.I. companies. But it may do little to ease concerns that the administration is intent on taking a more active role in the release of new A.I. systems.
Under the deal, Anthropic agreed to place safeguards on its A.I. models in exchange for restoring access to hundreds of clients. The Commerce Department had abruptly told the company on June 12 that it had 90 minutes to disable Mythos and Fable.
Anthropic later discovered that the administration’s concerns stemmed from a paper published by Amazon researchers that pointed to a perceived security shortcoming. The researchers were able to persuade the A.I. model to disclose flaws in bits of vulnerable software code.
Amazon did not include tests of models from other A.I. firms, some of which are capable of generating the same information, cybersecurity experts who had seen or been briefed on the report said.
It was unclear if the administration would fully lift the restrictions on Anthropic, or create a more formal a process that other A.I. companies could follow to avoid similar problems.
Anthropic and the Trump administration first came to blows at the start of the year over a $200 million Pentagon contract for A.I.
The company and the Pentagon disagreed over how the technology should be used in warfare. That led Defense Secretary Pete Hegseth to label Anthropic a “supply chain risk” in February. The designation, which meant the company posed a risk to national security, had never been used against an American company. In March, Anthropic sued the government over the label.
Anthropic released Mythos in April to a small group of companies and researchers. The A.I. company said its new A.I. model was particularly effective at finding vulnerabilities in software, raising concerns among security researchers and government officials that it could be used in cyberattacks.
Anthropic followed up with Fable, which it said had so-called guardrails that would limit what users could do with it. But researchers speculated that those guardrails could easily be broken, as was the case with most other A.I. systems.
When the government ordered Anthropic to take its new models offline, some at the company questioned whether the company was being unfairly singled out. But the administration’s interest in managing the release of A.I. models appears to be expanding.
Earlier on Friday, OpenAI, Anthropic’s main rival, unveiled an A.I. technology called GPT-5.6 Sol, saying it matched the performance of Mythos on standard benchmark tests used by A.I. companies. OpenAI said it would initially share the technology with only a small number of companies approved by the administration.
“We don’t believe this kind of government access process should become the long-term default,” the company said in a blog post. “It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them. We are taking this short-term step because we believe it is the strongest path to broader availability.”
(The New York Times has sued OpenAI and Microsoft, claiming copyright infringement of news content related to A.I. systems. The two companies have denied those claims.)
The administration also pressed Meta this week to submit its A.I. models for voluntary review.
Meta is the only major U.S. developer of A.I. technology that has not reached an agreement to voluntarily share its models with the federal government for review. OpenAI, Anthropic, Google, xAI and Microsoft all agreed to submit their models to the government’s A.I. safety group, known as the Center for A.I. Standards and Innovation.
Overseen by Mr. Lutnick, the center was created by the Biden administration to vet A.I. models and has a technical staff to lead those evaluations.
Cade Metz, Tripp Mickle and Tyler Pager contributed reporting.
