OpenAI rolled out a limited preview of its GPT-5.6 series—Sol, Terra and Luna—to a handful of U.S. customers vetted by the federal government. The move follows a Trump administration export‑control order that forced Anthropic to pull its Fable model after a security flaw was reported. By restricting access, OpenAI aims to keep the most powerful tools out of hostile hands.

Sol, the flagship of the new line, is billed as the most advanced model for cyber‑security tasks. OpenAI says it adds layered safeguards—training filters, real‑time response checks and account reviews—to block jailbreak attempts and refuse prohibited actions. The company also claims Sol can locate and remediate vulnerabilities more effectively than previous versions, while limiting its ability to launch end‑to‑end attacks.

Only U.S.-based, government‑approved entities have received the models so far, but OpenAI hints at expanding access to international partners as early as next week. The short‑term restriction lets the firm work with regulators on a formal release framework mandated by President Trump’s recent executive order. For investors, the episode underscores growing policy pressure on AI firms to police powerful capabilities before broad commercialization.