OpenAI has introduced a limited preview of its next-generation GPT-5.6 models, giving early access to a select group of trusted developers while broader public availability remains on hold during a U.S. government security review.

The preview includes three new AI models, Sol, Terra and Luna, each designed for different workloads ranging from advanced reasoning and software development to faster, lower-cost AI applications. OpenAI said the phased rollout is intended to gather feedback and validate safety measures before expanding access to a wider audience.

Among the new releases, Sol has emerged as the flagship model. According to benchmark results released alongside the preview, the model achieved leading performance on TerminalBench 2.1, a widely watched evaluation for terminal-based coding and problem-solving tasks. The results position Sol among the strongest AI systems currently available for software engineering and technical reasoning.

The company said early access has been restricted to selected API customers and Codex partners. The limited rollout follows discussions with U.S. government officials regarding the capabilities of advanced AI systems, particularly in areas related to cybersecurity and autonomous software development.

OpenAI stated that additional safeguards have been integrated into GPT-5.6 before its release. These include enhanced monitoring systems designed to detect potentially harmful or unauthorized use, along with improved oversight mechanisms intended to reduce security risks while maintaining model performance.

The decision to delay a wider launch reflects increasing government attention on highly capable AI models. As generative AI systems become more advanced, regulators and national security agencies are paying closer attention to technologies that could influence cybersecurity, critical infrastructure and other sensitive sectors.

The announcement has generated mixed reactions across the technology community. Many developers welcomed the reported improvements in coding, scientific reasoning and technical accuracy, viewing GPT-5.6 as another significant step forward in AI capabilities.

At the same time, the limited availability has sparked discussion about how advanced AI systems should be released. Some developers have questioned whether restricting access to selected organizations could slow innovation or create unequal access to emerging technologies during an increasingly competitive AI race.

The launch also highlights a broader shift in how frontier AI models are being introduced. Rather than releasing increasingly powerful systems to the public immediately, developers are adopting phased deployment strategies that combine technical testing with safety reviews before expanding access.

OpenAI has indicated that GPT-5.6 will become more broadly available after the preview period concludes and additional evaluations are completed. The company has not announced a specific timeline for the wider rollout but said it plans to expand access once the necessary security assessments and testing processes have been finalized.

The preview marks another milestone in the rapid evolution of artificial intelligence, while also underscoring the growing role governments are playing in shaping how the most advanced AI technologies reach developers and businesses.