OpenAI delays GPT-5.6 after government request

The concrete change for anyone planning to deploy on frontier models is that availability can now depend on a government security-review clock the vendor does not fully control. OpenAI is releasing GPT-5.6 as a limited preview to select enterprise partners rather than broadly, and the government will approve access customer by customer during that window. That combination of a technical gate (limited preview) and a policy gate (per-customer approvals) slows broad availability, concentrates early usage data in a small partner cohort, and adds compliance overhead to access management. Teams building on GPT-5.6 should plan procurement timelines around review, not just model readiness.

The Trump administration asked OpenAI to stagger the rollout of its next major model over security concerns, The Information first reported, with Reuters, Bloomberg, The Verge, and Axios describing it as the first time the US government has preemptively asked an American AI company to restrict a commercial launch before release. Reuters reports the request came from the Office of the National Cyber Director and the Office of Science and Technology Policy, and that CEO Sam Altman told employees the government would approve preview access customer by customer.

Both OpenAI and the administration view GPT-5.6 as capability-equivalent to Anthropic's Mythos, particularly in cybersecurity-relevant abilities, per Axios, and that parity triggered the same staggered-access framework applied earlier to Anthropic's Mythos 5 and Fable 5. The signal is that gating attaches to a capability threshold, not to one company, so any lab crossing that bar could face similar pre-launch review. Whether this hardens into a formal policy framework or stays an ad hoc bilateral arrangement is the open question.

Track whether the government-approval mechanism is codified, which enterprise partners get early access, whether non-US customers are treated differently, and whether comparable requests precede the next frontier launch. Each would confirm this as a durable deployment constraint rather than a one-off.