OpenAI has unveiled the new GPT-5.6 family of frontier AI models, consisting of three variants designed for different enterprise workloads: Sol, Terra, and Luna. The announcement comes amid a delicate geopolitical context, with the U.S. government requesting a phased rollout. Initially, only about twenty selected organizations will gain access, pending completion of security evaluations mandated by an executive order signed by President Trump earlier this month.
Sol, Terra and Luna: Three AI Profiles for Three Types of Workloads
GPT-5.6 Sol is the flagship model, built for the hardest tasks such as advanced coding, cybersecurity research, and deep reasoning. It delivers top-tier performance but at the highest cost: $5 per million input tokens and $30 per million output tokens. GPT-5.6 Terra sits in the mid-range, balancing power and efficiency for high-volume production environments like customer support and document analysis, priced at $2.50 and $15 per million tokens. GPT-5.6 Luna is the lightest and fastest option, ideal for summarization, drafting, and routine automation, at $1 and $6 per million tokens. OpenAI has moved away from the 'nano' and 'mini' naming scheme, adopting celestial names to emphasize distinct capability tiers rather than scaled-down versions.
Sponsored Protocol
Benchmark Records and New Reasoning Modes
Tests show a clear improvement over the GPT-5.5 series. On TerminalBench 2.1, which evaluates command-line automation, Sol achieved a record 91.91% using the new 'ultra' mode, surpassing GPT-5.5's 83.4% and Claude Mythos 5's 88%. On Agent's Last Exam, Sol is the only model to cross the 50% completion threshold in code mode. Technical innovations include a 'max reasoning' setting for problems requiring extended deliberation and the introduction of sub-agents that decompose and accelerate complex projects.
Sponsored Protocol
Security and Compliance Implications for Enterprises
OpenAI classifies all three models at 'High' risk for cyber and biological-chemical capabilities, imposing potential governance obligations even for the cheaper tiers. The company dedicated approximately 700,000 A100e GPU hours to automated red-teaming, implementing a multi-layered safeguard stack including model-level refusals, real-time screening, and activation classifiers. For Sol and Terra, these classifiers can pause output streaming if risky patterns are detected. OpenAI acknowledges possible false positives, with 81.6% recall on cybersecurity evaluations. Companies using these models for defensive purposes may face unintended blocks.
Geopolitics of AI: Between Government Control and OpenAI's Criticism
The limited release is a direct result of coordination with the White House, which requested a 30-day window to assess model safety. OpenAI has publicly criticized this approach, stating it should not become the long-term default as it keeps the best tools away from users and developers. The move follows U.S. restrictions on Anthropic, which led to the removal of Claude Fable 5 and Mythos 5 models. For European enterprises, as highlighted in a previous analysis on OpenAI and Italian SMEs, this situation may worsen the AI access gap. Additionally, the arrival of Paul Meade, former head of Apple's AI smart glasses, at OpenAI signals a talent acquisition strategy, as detailed in this article.
Sponsored Protocol
For more on retrieval architectures and indexing, see the Wikipedia page on GPT-4.
