Sam Altman informed staff that the United States government wants OpenAI to release its upcoming model, GPT-5.6, initially to a small list of trusted partners, with access approved on a case-by-case basis for each customer. For years, the debate about slowing down powerful AI models has been largely a concern for companies’ security teams and outside critics.
See also: IBM joins forces with OpenAI to integrate AI into enterprise security
Now, a government request has been added. The Trump administration has asked OpenAI to delay the release of its upcoming model, marking the first time the US government has preemptively asked an American AI company to curtail a release before it happens.
Altman told employees that the government has asked the company to initially release the model to a select group of trusted partners before a broader rollout. The government will approve client-to-client access during this preview period. This request came from discussions with two government agencies: the Office of the National Cybersecurity Director and the Office of Science and Technology Policy.
The concern is framed as a cybersecurity issue, not a competition or content issue. The concern is what a fairly capable model could do in the wrong hands, and the phased rollout is intended to limit that exposure during the initial phase. The timing of this request aligns with a broader shift. It comes about two weeks after rival Anthropic’s most capable offerings were pulled from the market under government order, suggesting that Washington is now actively influencing the release schedules of leading AI labs rather than simply reacting to them.
See also: Google Gemini co-head Noam Shazeer leaves for OpenAI
The proposed mechanism is significant. A client-by-client approval process during a preview period would give a government agency, if implemented as described, direct control over who gets early access to a pioneering model. This approach mirrors the phased release OpenAI used for GPT-5.4-Cyber, which was released to vetted security groups under a Trusted Access.
This represents a more proactive approach than the voluntary commitments and ex post evaluations that have characterized U.S. AI policy to date, shifting control of a release, at least temporarily, from the company to the state. For OpenAI, this arrangement has multiple implications. A gradual release slows the company’s ability to introduce its newest model to paying customers and developers just months after GPT-5.5 was released to the enterprise market, which could have commercial implications in a fast-moving market.
It also provides a degree of political cover; a model released with explicit government involvement is harder to pin down on the company if problems arise.
How OpenAI will balance these factors will become clearer as the preview period and subsequent developments unfold. Many of the details are based on Altman’s account to staff and reports from sources rather than an official government statement, and OpenAI has not published the terms of the settlement. The model name, client-to-client approval mechanism, and the services involved come from those accounts.
See also: Beware! New attack against OpenAI Codex users
This episode, if confirmed, establishes a new stance: a US administration treating the release of a breakthrough model as something that should be limited, with a leading laboratory agreeing to this approach. The next question is whether this will become the standard for future releases.
