Agentic AI,
Artificial Intelligence & Machine Learning,
Governance & Risk Management
Enterprise Users Face Delayed Access as Trusted Partners Get Early Preview
In a significant development in the sector of artificial intelligence and machine learning, OpenAI has decided to limit the rollout of its latest model, GPT-5.6. This decision, made in response to a request from the Trump administration, restricts access to a select group of users, delaying broader availability for enterprise users. The model is known for its strong capabilities in coding, research and development, as well as cybersecurity.
This strategic decision to stagger the rollout comes just weeks after the government imposed emergency export controls on Anthropic’s Fable 5 model, citing security concerns. The Trump administration highlighted vulnerabilities in the Fable model, expressing concerns that cyber adversaries could exploit these weaknesses to engage in harmful activities that pose national security risks. On June 5, in a parallel move, Trump issued an executive order directing frontier labs to voluntarily submit their artificial intelligence models for safety evaluations.
Reports indicate that OpenAI’s CEO, Sam Altman, communicated to employees that the administration specifically requested the release of the GPT-5.6 model to a limited selection of partners prior to its full launch. This approach has been confirmed by OpenAI in a blog post detailing the model’s release. The company has stated that the new model will first be accessible to a “limited preview” group, consisting of trusted partners whose involvement has been pre-approved by the government.
While there has been speculation surrounding the identities of these trusted partners, OpenAI has not responded to inquiries seeking further details. This strategy mirrors the earlier rollout of Anthropic’s Claude Mythos model, which also limited access due to concerns that the model could be misused if unleashed broadly.
The executive order allows for a voluntary 30-day review period for frontier AI models but does not specifically address the approval of user access. This omission leaves many enterprises uncertain about when they might gain access to the models they are eager to utilize. OpenAI claims that while it intends to adhere to government requests, it also aims to make the three tiers of GPT-5.6—Sol, Terra, and Luna—generally accessible in the coming weeks.
During this preview period, OpenAI plans to collaborate closely with its partners to continue testing the model while working towards a wider release. The company has expressed concerns that the current government access procedure should not become a long-term practice. OpenAI fears that such constraints may hinder the ability of users, developers, enterprises, and cyber defenders to access the powerful tools they need.
OpenAI underscores its belief that this limited rollout is the best route to achieving broader availability soon. The features of GPT-5.6 are impressive, with the flagship Sol Ultra version designed to assist in identifying vulnerabilities, understanding potential attack pathways, and developing effective solutions to counter threats. Additionally, Sol Ultra offers advanced coding capabilities, long-term reasoning, and orchestration for sub-agents, thus supporting large-scale research and engineering undertakings.
Industry experts have weighed in on the situation, highlighting the complexities surrounding the regulation of AI models. Jared Shulman, co-founder and CEO of fintech startup Daylit, remarked that the government’s intervention indicates the immense power wielded by models like GPT-5.6, though he recognizes that the current regulatory landscape appears reactive rather than proactive. Enterprises are increasingly opting for a model-agnostic approach since different models excel in varying tasks.
Shulman further explained that providing a small number of users exclusive access to GPT-5.6 creates a unique market dynamic, where organizations may feel pressured to secure their position on the privileged access list. This staggered approach establishes a new set of incentives within the industry.
Contrast Security’s co-founder and chief technology officer, Jeff Williams, added that the intersection of risk and innovation poses a dilemma for both the government and frontier labs. He noted that while this intervention provides a momentary pause to prepare for potential international threats, it does not really diminish the capabilities of the technology. Instead, it simply offers a slight delay in the race for technological advancement.
Moreover, many enterprises are now focusing on developing infrastructures and skills that can be employed by AI agents, often sidestepping the limitations presented by robust models like GPT-5.6. During the initial user preview, OpenAI intends to engage with enterprise customers on establishing long-term strategies, potentially involving features like privacy-preserving detection measures and customer-operated safety controls.
Shulman voiced concerns that adding conditions for identity or permission-based access to models could complicate matters, particularly at the application layer. The question of who is using the model and how it interacts with other AI agents remains an intricate challenge for enterprises.