OpenAI’s next model release is reportedly taking a different path. Instead of an immediate public launch, GPT 5.6 is expected to begin with a restricted preview for a limited set of close partners, reflecting a sharper focus on AI safety, cyber risk, and federal oversight.
The reported shift places OpenAI in the middle of a larger debate about how frontier AI models should be introduced when they may have powerful cyber capabilities. The central question is not only when a model should reach users, but who should decide which customers get early access.
A smaller first release for GPT 5.6
According to The Information, OpenAI does not plan to distribute GPT 5.6 broadly at the start. The company instead plans to share it only with a select group of close partners because the Trump administration told it to.
At a meeting this week, CEO Sam Altman reportedly told staff that the government would be “approving access customer by customer” during a preview period. He reportedly added that, if the restricted rollout goes well, OpenAI hopes to move to a general, broader release a “couple of weeks later.”
That would make the GPT 5.6 launch unlike OpenAI’s previous releases, at least in the way access is managed at the beginning. The early phase would not simply be a company-run preview. Based on the report, federal officials would have a role in deciding which customers can use the model during that window.
Why federal officials are involved
The reported request came from the Office of the National Cyber Director and the Office of Science and Technology Policy. The Information also reported that administration staffers “worked closely” with the government on the upcoming release.
That involvement is notable because the Trump administration had originally positioned itself as taking a “hands-off” approach to AI. More recently, however, it has pushed for federal oversight of new models. Earlier this month, Trump signed an executive order directing certain AI companies to voluntarily submit new models to the government for testing and evaluation before releasing them publicly.
The OpenAI situation appears to fit that changing posture. The government is not only evaluating models in the abstract; it is reportedly shaping how access to a major new model is granted before a wider launch.
The Anthropic comparison
The approach reportedly being asked of OpenAI resembles what Anthropic has already chosen to do voluntarily. Earlier this year, Anthropic announced that its new frontier cyber model, Claude Mythos, would only be released to a small coterie of partners through a program called Project Glasswing.
Anthropic said the model was too powerful to release widely and could cause more harm than good if misused. That argument led to debate. Some observers questioned whether the warning was marketing rhetoric, while others treated it as a serious attempt to limit access to a capable model. The source suggests the answer may be somewhere in between.
The comparison matters because it shows two different paths toward restricted AI access. Anthropic framed its limits as a voluntary safety decision. OpenAI’s GPT 5.6 rollout, as reported, is being limited after pressure from the Trump administration.
The cyber risk behind the caution
The concern is not simply that new AI models are more capable in a general sense. The specific worry centers on frontier cyber tools and their potential to identify and exploit software vulnerabilities quickly.
Cybercriminals have used automated tools for a very long time. Generative AI adds another layer because large language models have shown they can write malware, and some can even execute entire ransomware attacks autonomously.
The risk becomes sharper when a model can work through cyber tasks at speeds no human analyst could match. If a model can find hidden bugs and exploit them, it could create serious problems for organizations that depend on complex software infrastructure.
Those hidden bugs can serve as entry points into enterprise networks. For companies running large systems, that makes model access a practical security issue, not only a policy debate.
What remains unclear
The biggest uncertainty is how dangerous these closed models really are. Because systems like Claude Mythos are not publicly available, outside observers cannot easily test the strongest claims about their capabilities.
That uncertainty cuts both ways. Limited access may reduce the chance that a powerful tool is misused, but it also makes independent assessment harder. Companies, government offices, and selected partners may have more information than the broader public about what these models can actually do.
For OpenAI, the reported GPT 5.6 plan creates a narrow bridge between model development and public release. A small partner preview, government-reviewed access, and a possible broader launch a “couple of weeks later” all point to a more controlled release process for advanced AI.
For the AI industry, the message is broader: frontier model launches are becoming less like simple product updates and more like managed safety events, especially when cyber capabilities are part of the conversation.