US Launches Pre-Release Review for Frontier AI Models — GPT-5.6 and Claude Are First
Under Executive Order 14409, the US government has actually reviewed OpenAI's GPT-5.6 and Anthropic's Claude models. The process is voluntary, but access can still be restricted over national security concerns.
Following Executive Order 14409 (“Promoting Advanced Artificial Intelligence Innovation and Security”), signed on June 2, 2026, a real pre-release review process for frontier AI models has begun operating in the United States. OpenAI’s GPT-5.6 had its access restricted for 13 days before release, and Anthropic’s Claude Fable 5 and Mythos 5 had their global service fully suspended under a Commerce Department export-control order — citing the technical difficulty of real-time nationality verification — before being restored on June 30. These are the first real cases of a distinctly American approach to AI regulation actually playing out, in contrast with the EU’s delay of its high-risk AI rules.
The core of EO 14409
The executive order covers three broad areas:
- Strengthening the cyber defenses of government information systems
- Establishing a voluntary framework for frontier AI developers to engage with the government before release
- Expanding criminal enforcement resources against AI misuse
According to the Congressional Research Service (CRS), agencies including the NSA, CISA, and NIST must develop a classified process for determining which models qualify as “covered frontier models,” and participating companies may give the federal government access to a covered model for up to 30 days before its wider release.
Voluntary in name, different in practice
The White House’s own text explicitly states this is not a licensing or preclearance regime — companies can, in principle, decline to participate. But the real-world cases tell a different story.
| Company/Model | What actually happened | Basis |
|---|---|---|
| OpenAI GPT-5.6 | Access restricted for 13 days before release (limited to government and trusted partners) | Cybersecurity risk review by the Commerce Department’s CAISI (Center for AI Standards and Innovation) |
| Anthropic Claude Fable 5 / Mythos 5 | Global service fully suspended (June 12) over inability to verify user nationality in real time, restored June 30 | Commerce Department export-control order, following reports that a jailbreak could bypass the models’ safety guardrails |
Why it matters
US AI regulation has historically been seen as relatively light-touch compared to codified regimes like the EU AI Act. But the GPT-5.6 and Claude cases are the first to show that even a “voluntary” framework can, in practice, result in real access restrictions on national security grounds. Future frontier-tier model releases — including next-generation Claude and GPT series models — could plausibly go through the same process.
Limitations and caveats
- The detailed procedures of EO 14409 — including exactly how a model qualifies as “covered” — are classified, making the full process hard to verify externally.
- Claude Fable 5 and Mythos 5, covered in this article, are Anthropic model names this site hasn’t covered in a dedicated article before; their exact specs and release timing would need separate verification going forward.
- US AI regulatory posture can shift with changes in administration, so this article should be read as a snapshot as of September 2026.
Conclusion
EO 14409 was designed as a voluntary framework with no legal force, but GPT-5.6’s 13-day access restriction and the full suspension of Claude’s service showed that the US government can still exert real influence over when frontier AI models launch. It’s a different approach from the EU’s codified regulations, but both regions appear to be converging on the same basic idea: government review before a frontier model ships.