Anthropic brings Mythos 5 back for critical infrastructure

Anthropic is restoring access to Claude Mythos 5 for a limited set of U.S. organizations after negotiating partial relief from government restrictions on its newest models. Business Insider reported on June 27 that the U.S. government has cleared Mythos 5 for organizations involved in operating and defending critical infrastructure. Anthropic said the redeployment will happen quickly for those customers, while Claude Fable 5 remains unavailable.

The decision follows a sharp intervention earlier in June, when Anthropic said a U.S. export-control order forced it to disable access to both Fable 5 and Mythos 5. The order barred access by foreign nationals, including foreign national employees inside Anthropic. Because that rule was difficult to enforce cleanly, the company took the models offline more broadly while it worked with officials.

This new carveout matters because Mythos 5 is not a general chatbot upgrade. Anthropic has positioned it as its strongest cybersecurity model, aimed at sensitive work such as vulnerability analysis, defensive operations and critical systems. Letting some U.S.-based organizations use it again suggests that the government sees a difference between uncontrolled model access and access by trusted customers working on security-relevant infrastructure.

For AI users, the immediate effect is limited. Most people will not suddenly see Mythos 5 in Claude. For makers and companies, the signal is more important: frontier AI access is becoming more conditional. Model availability can now depend not only on price, subscriptions and technical capacity, but also on nationality, sector, customer vetting and the government's view of national-security risk.

The market impact is also broader than Anthropic. OpenAI said on Friday that the U.S. government had asked it to limit preview access to GPT-5.6 to a small group of trusted partners. Together, the two cases show a new pattern for the most capable models: release first to selected users, negotiate with policymakers, then widen access only if safety and political concerns are manageable.

That may frustrate developers who want immediate access to the best tools. But it also shows how quickly AI models have moved into the same strategic category as chips, cloud capacity and cyber capabilities. The question is no longer just which lab has the strongest model. It is who is allowed to use it, for what purpose, and under whose oversight.