Currently, access is restricted to a select group of trusted partners sanctioned by the US government, with plans for a broader rollout anticipated in the upcoming weeks.
The limited introduction followed a request from the Donald Trump administration to temporarily limit access to GPT 5.6, citing concerns over its sophisticated capabilities. While OpenAI acquiesced to this temporary measure, the company emphasized that it does not endorse a government approval process as the standard for accessing new AI models.
OpenAI affirmed this was a transitory arrangement and stated that it is collaborating with the US government to establish an improved long-term framework for future model releases. The organization noted that restricting its most advanced tools to a select few is not the ideal approach, as developers, businesses, researchers, security professionals, and other users require access as well.
In a statement, OpenAI remarked, “As part of our ongoing engagement with the US government, we previewed our plans and the models’ capabilities prior to today’s launch. At their request, we are initiating with a limited preview for a small circle of trusted partners whose involvement has been communicated to the government, before expanding availability. During this preview, we will persist in testing and closely coordinating with partners as we progress toward broader access.”
“We do not believe this sort of government access procedure should become the long-term norm. It deprives users, developers, enterprises, cyber defenders, and global partners of critical tools. We are undertaking this short-term measure because we believe it represents the optimal path to wider availability in the coming weeks, while we collaborate with the Administration to formulate the cyber Executive Order framework and a systematic procedure for future model releases.”
OpenAI stated that Sol is its most potent AI model to date, excelling in fields such as computer programming, biology, and cybersecurity. Furthermore, OpenAI noted that Sol outperformed Anthropic’s Mythos model on Terminal Bench 2.1, a benchmark used to assess AI systems’ capacity to tackle complex tasks.
Similarly, Terra and Luna have been developed with the most robust safety protocols implemented to date. The company clarified that each model possesses safety configurations tailored to its capability level.
As its AI models advance, OpenAI mentioned that it continuously enhances these protections to better manage misuse and real-world threats.
Concurrently, the organization emphasized the importance of ensuring that users can still utilize the models for legitimate purposes, including code reviews, identifying security vulnerabilities, debugging software, implementing updates, learning about cybersecurity, and testing systems to strengthen their security.
