OpenAI has announced three new GPT-5.6 models with significant improvements in safety, coding, and complex reasoning, but they're currently available only to a restricted group of "trusted partnerships and organizations" due to U.S. government requests. The company says it expects to release these models to the broader public within the coming weeks, though the government approval process has sparked internal frustration at OpenAI about the future of AI development.

What Are the Three New GPT-5.6 Models?

OpenAI introduced three models in the GPT-5.6 series, each designed for different use cases:

• Sol: The flagship model optimized for deep thinking and complex agentic work, where AI systems autonomously perform multi-step tasks. OpenAI claims Sol matches Anthropic's Mythos model for cybersecurity tasks while using a third of the output tokens, meaning it delivers similar performance with greater efficiency.
• Terra: A mid-tier model designed for everyday work and general-purpose tasks, balancing capability with cost-effectiveness for standard business applications.
• Luna: The smallest, fastest, and cheapest option in the series, intended for users and developers who prioritize speed and affordability over maximum capability.

All three models set a new standard for safeguards and protection against "adversarial pressure," according to OpenAI's announcement. They are designed to resist attempts to disguise user intent and to prevent jailbreaking, where users try to bypass safety restrictions through clever prompting.

Why Is the U.S. Government Blocking Access?

The restricted rollout stems from requests by the U.S. government, which wants to conduct its own testing and evaluation before these models become widely available. OpenAI is complying with those requests this time, but the company made its displeasure clear in its announcement.

"We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them," OpenAI stated.

OpenAI, in official announcement

This tension reflects a broader shift in how frontier AI models are released. Rather than launching directly to the public, advanced AI systems now face government scrutiny before widespread deployment. OpenAI's reluctance suggests the company views this as a temporary measure, not a permanent policy direction.

How to Prepare for GPT-5.6 Access When It Becomes Available

If you're interested in using these new models once they launch publicly, here are practical steps to consider:

• Monitor OpenAI's Official Channels: Follow OpenAI's website, blog, and social media accounts for announcements about public availability. The company has indicated a release "in the coming weeks," so staying informed will help you access the models as soon as they become available.
• Evaluate Your Use Case: Consider whether Sol, Terra, or Luna best fits your needs. Sol excels at complex reasoning and cybersecurity tasks, Terra handles everyday work, and Luna prioritizes speed and cost. Matching your use case to the right model will maximize value.
• Prepare for Potential Restrictions: Early testers may encounter "apparently unnecessary blocks and restrictions" as OpenAI refines safety features. Feedback from these testers will shape the final release, so expect some friction during the initial rollout phase.

What Performance Improvements Should Users Expect?

OpenAI claims the GPT-5.6 models deliver performance "competitive" with the previous GPT-5.5 generation, but with notable improvements across several dimensions. These include better affordability, enhanced safety mechanisms, improved agentic capabilities (the ability to autonomously complete multi-step tasks), stronger coding performance, advances in biology understanding, and superior cybersecurity capabilities. The company has not released detailed benchmark scores, but the emphasis on cybersecurity and coding suggests these models are being positioned as tools for technical professionals and security teams.

The efficiency gains are particularly noteworthy. Sol's ability to match Anthropic's Mythos model for cybersecurity while using a third of the output tokens means faster response times and lower computational costs for users running these models at scale. This matters for enterprises and developers who pay based on token usage.

What Does This Mean for the Future of AI Access?

The restricted rollout has already sparked concern among AI enthusiasts and developers. On Reddit and other forums, users have expressed resignation about the changing landscape of AI access. One user commented, "The days of the public getting access to these frontier models is gone," while another observed, "the divide has started" between those with early access and everyone else.

OpenAI's public pushback against the government approval process suggests the company wants to avoid setting a precedent where every new model requires government vetting before release. However, the fact that OpenAI complied with this request indicates the company has limited leverage to resist such demands. The coming weeks will reveal whether the government's testing process accelerates or delays the public rollout, and whether this becomes a standard practice for future AI releases.