OpenAI's staggered GPT-5.6 release follows a U.S. government request for pre-deployment review, setting a new precedent for AI model governance.
OpenAI released three new models on Friday but kept them out of reach for almost everyone. The company announced GPT-5.6 Sol, Terra and Luna, with Sol described as its strongest system yet, then immediately limited access to a "small group of trusted partners" at the request of the U.S. government The Guardian.
The move mirrors a pattern set weeks earlier by Anthropic. That company pulled its Mythos model entirely after the administration ordered it to block foreign nationals from accessing public versions with advanced cyber capabilities CNBC. OpenAI's blog post framed the restriction as a "short-term step" toward broader availability in coming weeks.
Government involvement has escalated since President Trump signed an artificial intelligence executive order earlier this month. The order asked developers to voluntarily submit models for capability assessment before full release. OpenAI said it previewed GPT-5.6's plans and shared partner names with officials ahead of Friday's launch CNBC.
The company's dissatisfaction was explicit. "We don't believe this kind of government access process should become the long-term default," OpenAI wrote, arguing it keeps the best tools from users, developers, enterprises, cyber defenders and global partners. Still, it said compliance represents the "strongest path to broader availability."
GPT-5.6 Sol, the flagship tier, shows improvements in coding and biology benchmarks according to the company. Terra and Luna occupy lower capability rungs. The naming scheme replaces the single-version numbering of prior generations. GPT-4.5 was simultaneously retired in ChatGPT, with existing conversations migrated to GPT-5.5 Releasebot.
The market reaction
Anthropic's experience suggests the precedent may harden. After initially delaying Mythos voluntarily, the company faced a formal export control directive that forced a complete shutdown of public access. Negotiations with Washington continue with no timeline for restoration. OpenAI appears to be negotiating a framework to make future reviews repeatable rather than ad hoc.
For developers and enterprises, the immediate effect is uncertainty. Access to frontier models now depends on government clearance processes that remain opaque. The administration has not published criteria for trusted-partner status or defined which capabilities trigger review. OpenAI did not disclose which organizations received early access.
The tension between national security and innovation velocity is now operational. If pre-deployment review becomes standard, release cycles will lengthen and smaller players without government relationships may fall further behind. The White House has not indicated whether the voluntary framework will become mandatory.
Next week's negotiations between Anthropic and federal officials may signal whether the trusted-partner model expands or collapses. Until then, the most advanced artificial intelligence systems sit behind a new kind of gatekeeper.
