The US lifted its block on Anthropic's Mythos 5 for a named list of companies

The US government lifted its block on Anthropic's Claude Mythos 5 on Friday, ending a roughly two-week freeze that had pulled the company's most capable model off the market. The decision came in a letter from Commerce Secretary Howard Lutnick to Anthropic's chief compute officer, Tom Brown, and it clears the model for release to more than 100 US institutions, including large companies and government agencies, according to Semafor, which first reported the letter.

The mechanism matters here, because it is not the one the White House has been advertising. This ran on export-control law. Two weeks ago Commerce imposed export controls on Mythos, which shut the model down along with its cousin Fable 5, after Amazon and other companies warned the models could be jailbroken for malicious use. Lutnick's letter unwinds that by waiving the license requirement for a specific set of customers: "a license will no longer be required to export, reexport, or in-country transfer" the model "to entities identified in Annex A to this letter," the letter says, per Semafor. In other words, a named list of approved recipients, which is what an export license is.

"I have determined that appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model," Lutnick wrote, citing what he called significant progress in daily talks with the company since the block took effect. Commerce spokesman Benno Kass framed the speed as the point: "In just two weeks, we have worked diligently to ensure America remains the global leader in AI while safeguarding our security."

What the block was about

The trigger, as Semafor reported earlier this month, was concern that Mythos had reached partners too closely tied to China. The reporting pointed to a South Korean telecom, which Wired identified as SK Telecom. The jailbreak warnings from Amazon and others gave the action a security rationale; the China-access worry gave it urgency. The result was the same: a frontier model that was live one week was dark the next, by order of the Commerce Department.

Fable 5 is still dark. Lutnick's letter says nothing about it, and people close to the talks told Semafor they expect it to come back without putting a date on it. Fable was briefly the most powerful model a consumer could actually use, so its absence is the part most users will notice.

The same day, OpenAI shipped behind the same kind of gate

Hours before the Anthropic letter went out, OpenAI released GPT-5.6 and immediately limited it. The model comes in three versions, Sol, Terra, and Luna, with Sol the most capable, and OpenAI says it is available only to about 20 companies that the government individually approved, with broader access promised "in the coming weeks," per Axios.

OpenAI put its objection on the record, which is unusual for a company mid-negotiation with the people regulating it. "We don't believe this kind of government access process should become the long-term default," it wrote. "It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." It also said it was accepting the arrangement as "the strongest path to broader availability" while it helps the administration build "a repeatable process for future model releases." On the specific fear driving all of this, OpenAI argued GPT-5.6 Sol "is better at helping people find and fix vulnerabilities than reliably carrying out end-to-end attacks," and that the model does not hit the "critical" cyber threshold in its own preparedness framework.

Where the rules actually are

The framework everyone keeps pointing to is Executive Order 14409, "Promoting Advanced Artificial Intelligence Innovation and Security," signed June 2. It directs agencies to build a classified benchmarking process, due by roughly August, to decide which systems count as "covered frontier models," and it sets up a voluntary channel for developers to give the government up to 30 days of pre-release access to models with advanced cyber capabilities.

The order also says, in plain text, that nothing in it authorizes "a mandatory governmental licensing, preclearance, or permitting requirement" for AI models. That line is worth holding next to what just happened. The benchmarking process that would define a covered frontier model does not exist yet. What exists is export-control authority, which Commerce used to switch Mythos off and then, via a letter naming approved recipients, switch it back on for them. Whatever the executive order says about licensing, the tool that did the work this month was a license. Holland & Knight, in a client note on the order, called it a strategic shift for an administration that had until recently championed a hands-off approach, and Fortune described the Anthropic situation earlier in June as a licensing regime by another name.

What is measurable today is small. About 100-plus institutions can use Mythos 5; about 20 can use GPT-5.6 Sol; Fable 5 is offline with no date; and allied governments and overseas customers, as Semafor notes, are waiting on decisions made in Washington with no visibility into the criteria. The benchmark that is supposed to make any of this legible is still a few months out.