U.S. eases Anthropic AI restrictions, allows Claude Mythos 5 access

The Commerce Department allowed Anthropic to restore Claude Mythos 5 access to more than 100 approved U.S. institutions, a selective reversal that eases a two-week freeze without opening the model to the broader market. The new guidance says an export license is no longer required for Mythos 5 when the user is on the approved list, including certain foreign-national employees of those trusted organizations and Anthropic’s own foreign-national employees.

The decision marks a sharp de-escalation after the government forced Anthropic to disable access to Mythos 5 and Fable 5 on June 13, 2026, while it complied with export-control demands tied to foreign-national access. In a letter addressed to Anthropic co-founder and chief compute officer Tom Brown, Commerce Secretary Howard Lutnick said Anthropic had worked with the government to address the risks attached to the covered models and that the effort had produced significant progress.

The approval does not amount to a full restoration. Companies and agencies outside the approved list remain blocked, so access stays tightly gated rather than broadly commercial. The letter also left Fable 5 untouched, a notable omission because it means the government has not yet said whether Anthropic’s other frontier model will get the same treatment. A source familiar with the matter said officials were moving toward allowing Fable 5 as well, but no timing was set.

The policy shift is the latest turn in a fraught relationship between Anthropic and Washington. The Congressional Research Service has said the dispute followed President Donald Trump’s Feb. 27, 2026, order directing agencies to stop using Anthropic technology, and that the Defense Department later designated Anthropic a supply-chain risk. The core disagreement centered on Anthropic’s refusal to allow its models to be used for mass domestic surveillance or fully autonomous weapons systems.

Anthropic has argued that its systems are being used in mission-critical national security roles and has warned that Mythos carried hacking risks, which helped keep it from wide release. Its Project Glasswing program shows how the company has tried to balance that risk with controlled access: it began with about 50 partners and expanded in June to roughly 150 organizations in more than 15 countries, all of them required to meet security requirements before gaining access.

The partial rollback also sets a template for future frontier-AI controls: approved partners can keep working, but only inside a government-managed gate. On the same day, OpenAI also limited access to a new model rollout to a small group of government-approved partners, underscoring how access rules are becoming part of the U.S. playbook for advanced AI distribution.