OpenAI has begun a restricted preview of its latest frontier model family, GPT-5.6, introducing three specialized variants: Sol as the flagship powerhouse, Terra as a balanced mid-tier option, and Luna as the fastest and most economical choice. The rollout emphasizes meaningful gains in complex reasoning and cybersecurity while introducing elevated safety considerations that have shaped a cautious, phased release strategy.
Sol stands out as OpenAI’s most capable model to date, delivering advances in agentic coding, long-horizon planning, and professional workflows.
It sets a new high mark on Terminal-Bench 2.1, a demanding evaluation of command-line tasks that require iterative planning and tool use.
Terra matches or approaches the performance of prior generations at roughly half the cost, positioning it well for everyday enterprise and development work.
Luna prioritizes low latency and affordability, targeting high-volume applications where speed matters most.
All three models support new reasoning modes, including deeper “max” effort settings and an “ultra” configuration that coordinates sub-agents for especially intricate problems
.Cybersecurity capabilities represent one of the clearest areas of progress.
The GPT-5.6 family demonstrates stronger performance in vulnerability research, identification of exploitation primitives, and related security tasks across multiple internal and external benchmarks.
Models show particular strength in helping users locate and remediate weaknesses, often outperforming earlier versions in defensive scenarios.
However, these improvements come with new risk classifications: under OpenAI’s Preparedness Framework, Sol, Terra, and Luna are all rated as high-capability in both cybersecurity and biological/chemical domains.
While evaluations indicate the models remain below the “Critical” threshold for fully autonomous end-to-end attacks on hardened systems, the increased proficiency in long-horizon cyber tasks and dual-use knowledge has prompted heightened scrutiny.
To address these risks, OpenAI has deployed its most comprehensive safety architecture yet.
The system combines model-level training to refuse prohibited cyber assistance, real-time classifiers that can pause or escalate sensitive generations, and extensive red-teaming involving hundreds of thousands of GPU-equivalent hours plus human experts.
Account-level monitoring helps distinguish legitimate defensive work from persistent misuse, while layered safeguards aim to make offensive applications more difficult, uncertain, and detectable without overly restricting beneficial uses such as vulnerability patching or security research.
Access during the preview is limited to a small group of approximately 20 trusted partner organizations, made available through the OpenAI API and Codex after coordination with US government officials.
Broader availability across ChatGPT, Codex, and the API is expected in the coming weeks once additional testing and refinement are complete.
The introduction of the GPT-5.6 family illustrates the accelerating pace of AI capability gains alongside the growing complexity of managing associated risks. By prioritizing defensive strengths while implementing robust controls, OpenAI seeks to expand access responsibly as these digital tools move toward wider deployment.