US eases ban on two Anthropic models after just weeks

The US government has eased an export ban on two AI models from the company Anthropic. The ban came just days after Anthropic released the models, which are built on Mythos, in early June. According to NPR, the reason was that people had found ways to bypass the models' safeguards, the built-in barriers meant to prevent misuse.

From partial to full opening

Around June 27-28, the Commerce Department partially eased the ban. A limited list of American companies, including their foreign employees, got access to one of the two models, Mythos 5.

According to the Indian outlet Outlook Business, citing Reuters sources, the export controls were removed entirely on July 1, 2026. The sources give slightly different dates for the easing.

Commerce Secretary Howard Lutnick reportedly set three conditions for Anthropic, according to Outlook Business:

- The company must proactively detect and handle security risks. - It must work with the US government on protocols. - It must report malicious activity.

It is unclear exactly which security concerns were behind the order. Anthropic says the order did not specify them.

OpenAI agrees to screening

The case also affects rival OpenAI. According to NPR, the company agreed to let the government screen which companies get access to its newest model. OpenAI argued, however, that such screening should not be the norm in the long term.

According to Outlook Business, OpenAI delayed the full public launch of GPT-5.6 at the government's request and limited access to a small group of screened partners.

OpenAI CEO Sam Altman has said, according to Outlook Business: "Extensive safety testing is not a bad idea, but I don't like the idea of the government choosing the customers."

Why does it matter?

The models are considered extremely powerful. According to NPR, Cisco security chief Anthony Grieco said Mythos (the preview version) could scan 1.8 billion lines of code in eight weeks, work that would otherwise take humans and tools combined eight years or more. That does not apply to the two released, weakened models. For ordinary users, the case shows that new AI models can be delayed or limited to select customers if authorities fear misuse.