Thought Leadership
OpenAI has released a preview of GPT-5.6, a new flagship model generation. The Sol model is optimized for complex reasoning tasks and software vulnerability analysis.
OpenAI has announced a limited preview of its new GPT-5.6 model family, which is divided into three performance tiers: Sol, Terra, and Luna. The version number will now represent the model generation, while the names indicate permanent capability levels. The most powerful model, Sol, is designed for complex logical tasks. Terra targets everyday workloads at lower costs, while Luna serves as a fast and cost-efficient option.
Following consultations with the U.S. government, the release will initially be restricted to selected partners. The temporary limitation is intended to evaluate potential national security risks associated with advanced AI systems in line with current regulations.
GPT-5.6 Sol focuses on defensive cybersecurity tasks
In cybersecurity benchmarks on the ExploitBench platform, Sol achieved results comparable to competing systems such as Mythos Preview while requiring only one-third of the output tokens. During source code reviews of Chromium and Firefox, the model identified software flaws and underlying mechanisms but did not create fully functional attack chains.
OpenAI emphasizes that the model is better suited for defensive cybersecurity applications, including vulnerability detection and software patch development.
To reduce potential misuse risks, the infrastructure relies on a multi-layered security architecture. Automated real-time classifiers analyze prompts related to biology and cybersecurity. If suspicious activity is detected, generation is paused and the context is reviewed by a secondary model.
To protect against universal system bypass attempts, OpenAI conducted more than 700,000 GPU hours of automated security testing before release. A broader rollout for ChatGPT and the general API is planned for the coming weeks.
(ll)
