The U.S. government cleared Anthropic on Friday to switch its Mythos 5 model back on for a curated list of about 100 organizations, two weeks after export controls forced the company to pull it offline entirely. Commerce Secretary Howard Lutnick laid out the terms in a letter to Anthropic chief compute officer Tom Brown. Fable 5, the consumer-facing version, got no such reprieve.
Who actually gets in
This is not a general re-release. Access goes to organizations that operate and defend critical infrastructure, and the stated purpose is defensive cyber work, autonomous vulnerability hunting and patching. Many of the cleared names come from Anthropic's invite-only Project Glasswing program, which before the shutdown reportedly counted around 200 firms including the usual roster of large tech companies and banks.
So roughly half the original Glasswing group makes the new cut, if those numbers hold. Nobody outside the talks has published the list, and that opacity is already drawing complaints. A legislative counsel at the free-speech group FIRE, John Coleman, put it bluntly to one outlet: no one knows how the companies are picked or why everyone else is shut out. Hard to argue with that when the selection criteria live in a non-public annex.
The number nobody can fully verify
"Around 100" is doing a lot of work here. NBC and others put it at roughly 100 organizations citing people familiar with the matter. The Washington Post reported Anthropic had submitted a list of 111 organizations for early access, per two White House officials, which is where the figure probably traces back to. Whether all 111 made it through is unclear, and "agencies and companies" lumps federal users in with private ones without breaking down the split.
What Anthropic gave up to get here
Lutnick's framing was cautious. He determined that appropriate safeguards were now in place for trusted partners, citing significant progress in two weeks of daily negotiations. What Anthropic actually changed to satisfy the government's jailbreak concerns has not been disclosed, by either side.
The letter also notes Anthropic committed to working with the government on protocols and standards for future model releases. Read that as the real story: this episode is becoming the template for how frontier models reach the market, not a one-off fix. Lutnick kept discretionary power to amend the partner list or change license terms whenever he likes.
One operational win for Anthropic: the clearance restores access for non-American employees at approved organizations and at Anthropic itself, reversing the most disruptive piece of the June 12 order, which had barred any foreign national from touching the models.
Fable 5 is the harder problem
Mythos 5 access can be fenced off to about 100 named organizations with verifiable identities. Fable 5 went out to hundreds of millions of consumers. There is no clean way to vet that many users, which makes any conditional restoration genuinely messy to design. Anthropic confirmed in a public statement that talks continue, but with no clear path.
The same day Lutnick's letter went out, OpenAI limited its own GPT-5.6 launch to government-approved partners. Two labs, same pattern, same week. Anthropic and the administration were set to keep negotiating Fable 5 through the weekend.
