GPT-5.6 Access: Trusted Partners & U.S. Model Review

A launch limited to a partner list nobody could see

On June 26, 2026, OpenAI announced three new models, GPT-5.6 Sol, Terra and Luna, and said it was complying with a U.S. government request to limit their initial rollout to a "small group of trusted partners." OpenAI said it had previewed the models' capabilities and shared its release plans with the government beforehand, and that it did not disclose the names of the partners who could use them. [1][3]

OpenAI described the restriction as temporary, saying it expected broader availability, through ChatGPT, Codex and the API, within weeks, and quoted its own blog post objecting to the arrangement: "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." [1][3]

A voluntary review, not an order

The request traced to an executive order President Trump had signed earlier that month, asking AI developers to voluntarily submit their most advanced models for government review, on terms the order left largely undefined, for up to 30 days before release. OpenAI said it was working with the administration to turn that into a "repeatable process for future model releases." [1][3]

Atlas interpretation: "Voluntary" did most of the work in that sentence. Nothing forced OpenAI to accept a delayed, partner-only rollout of its flagship model, but declining an administration request days after a rival's models were pulled offline entirely was not a real option either. [3]

Not the same thing that happened to Anthropic

The announcement came two weeks after the government had ordered Anthropic to pull Fable 5 and Mythos 5, models it had released three days before that order, offline for every foreign national worldwide. Anthropic took both models down entirely rather than screen access in real time, and reporting at the time noted the government had also "effectively banned" Mythos 5 on the same terms. [3]

Atlas interpretation: The two companies did not end up in the same place. OpenAI kept GPT-5.6 live for a partner list it controlled, under a voluntary review it could in principle have refused; Anthropic's models were unavailable to anyone, under a binding directive citing no published legal basis. Describing both as "agreeing to vet customers" flattens a request one company accepted into an order the other one didn't get to negotiate. [3]

A former official's warning about where this goes

Dean Ball, a former White House AI adviser who had by then agreed to join OpenAI, argued the executive order had created what he called a de facto involuntary licensing regime for frontier AI: without published safety standards to satisfy, a company has no fixed bar to clear and no way to know when review ends. He warned that open-ended delays of this kind could cede ground to Chinese competitors and put AI infrastructure investment at risk. [3]