OpenAI has introduced a limited preview of its advanced GPT-5.6 model series, featuring Sol, Terra, and Luna. The flagship model, Sol, is hailed as OpenAI’s most secure AI model and is currently available to a select group of partners. This release follows a request from the U.S. government, emphasizing the need for a staggered rollout due to the model’s capabilities and national security considerations.
New Tiered Model System
The GPT-5.6 series introduces a new tiered naming convention with Sol, Terra, and Luna. Sol stands as the leading model, Terra offers balanced capabilities at a reduced cost, and Luna provides affordable AI solutions. This tiered approach allows each model to evolve independently based on its capabilities.
Pricing for these models is structured based on token usage. Sol is priced at $5 for input and $30 for output per million tokens. Terra costs $2.50 for input and $15 for output, while Luna is the most economical, priced at $1 for input and $6 for output. Sol will also be available on Cerebras, offering high-speed processing for enterprise customers starting in July.
Controlled Access and Government Involvement
The U.S. government, particularly the Office of the National Cyber Director and the Office of Science and Technology Policy, has influenced the controlled release of GPT-5.6. Commerce Secretary Howard Lutnick also advised against a full-scale launch. OpenAI CEO Sam Altman noted that the government will approve access on a case-by-case basis during the preview phase.
While OpenAI complies with current requirements, the company expressed its preference for a more open access model in the future. They believe that restricting access limits the availability of advanced tools to users, developers, and enterprises who could benefit from them.
Advanced Security Features
GPT-5.6 Sol is equipped with OpenAI’s most sophisticated security measures to date. It excels in tasks requiring long-term security solutions, enhancing efficiency without compromising performance. On ExploitBench, Sol competes with top models like Anthropic’s Mythos while using significantly fewer tokens.
The model also demonstrates superior performance on Terminal-Bench 2.1, achieving a score of 88.8%, and even higher in its ultra configuration. This configuration employs subagents to expedite complex workflows, outperforming previous benchmarks.
Sol incorporates multiple layers of cyberattack protection, including model-level safeguards, real-time misuse classifiers, and account-level reviews. These measures ensure that legitimate security research continues while minimizing risks associated with prohibited activities.
Despite its capabilities, GPT-5.6 Sol does not breach the Cyber Critical threshold under OpenAI’s Preparedness Framework, maintaining a focus on responsible use. It excels in identifying bugs and exploitation primitives, but does not autonomously generate complete exploits.
Looking Ahead
OpenAI’s staggered release strategy for GPT-5.6 highlights the importance of balancing innovation with security. The company continues to refine its models to provide cutting-edge AI solutions while adhering to national and international security standards.
This launch sets a precedent for future AI developments, emphasizing the need for comprehensive governance and responsible deployment. As the preview period progresses, broader access to GPT-5.6 is expected, paving the way for innovative applications across various sectors.
