According to The Information, OpenAI will not follow its previous public-release pattern for the newest model, GPT-5.6. Instead, the company plans to make the model available initially only to a small group of close partners after instructions from the Trump administration to restrict distribution during a preview period.
What OpenAI’s leadership said
The Information reports that Sam Altman, OpenAI’s CEO, told staff at a meeting this week that the government would be “approving access customer by customer” during the preview. Altman reportedly added that, if the limited release proceeds smoothly, OpenAI hopes to issue a broader public release “a couple of weeks later.”
Government involvement and agencies named
The report states that OpenAI staff “worked closely” with government officials on the upcoming release and that the agencies requesting a limited rollout included the Office of the National Cyber Director and the Office of Science and Technology Policy.
Those developments are part of a recent shift in the Trump administration’s posture toward AI. Although the administration initially framed its approach as relatively hands-off, it has in recent months pushed for federal oversight of new models. Earlier this month, President Donald Trump signed an executive order directing certain AI companies to voluntarily submit new models to the government for testing and evaluation before public release.
Parallels with Anthropic’s approach
The Information notes a similarity to a strategy already used by Anthropic. Earlier this year, Anthropic announced that its frontier cyber model, Claude Mythos, would be released only to a small set of partners via a program called Project Glasswing. Anthropic said the model was powerful enough that broader public release could cause harm in the wrong hands.
Observers have debated whether Anthropic’s public rationale reflects genuine safety concerns, marketing strategy, or a mix of both.
Why this matters: risks and uncertainty
The report underscores longstanding concerns about automated tools in the hands of cybercriminals, a risk that grows as generative AI becomes more capable. Large language models have been shown to assist in writing malware, and some systems may be capable of executing ransomware attacks autonomously.
Frontier cyber tools like Claude Mythos raise specific worries because they appear designed to find and exploit software vulnerabilities at speeds unattainable by human analysts. Given that many enterprise systems contain hidden bugs that can serve as entry points to networks, such capabilities could pose a significant threat to organizations running complex software infrastructures.
At the same time, the true scale of the risk is hard to assess publicly because these most advanced models remain closed to general access.
Summary
Per The Information, OpenAI will start GPT-5.6 with a restricted, partner-focused preview at the Trump administration’s request, with the government approving access on a customer-by-customer basis. If the limited release goes well, OpenAI may broaden availability after a few weeks. The episode highlights tensions among AI developers, public safety officials, and policymakers over how to balance access to powerful models with efforts to reduce misuse.
