OpenAI has released a limited preview of its GPT-5.6 series, with three model variants: Sol, Terra, and Luna. Due to a U.S. government security review framework for frontier AI models, the initial release is only available to a small set of trusted partners, but the company plans to roll it out to the public in the coming weeks.
The flagship model, Sol, introduces an Ultra mode that cranks up reasoning and enables sub-agent collaboration. On the Terminal-Bench 2.1 command-line workflow benchmark, Sol scored 88.8%, and in Ultra mode it hit 91.9% — beating GPT-5.5's 83.4% and Claude Fable 5's 88.0%. The mid-range Terra model approaches GPT-5.5's performance at half the price, while the lightweight Luna is designed for everyday automation tasks. Sol is priced at $5 per million input tokens and $30 per million output tokens, with prompt caching available to cut costs on repeated queries.
On the safety front, evaluations confirm Sol hasn't crossed any critical cybersecurity thresholds in the Preparedness Framework. OpenAI invested over 700,000 A100-equivalent GPU hours in automated red-teaming and equipped the entire model lineup with a defense stack that includes refusal mechanisms, real-time abuse classifiers, and account-level auditing. While this limited release follows the government's security framework, OpenAI stresses that it doesn't want government-controlled access to become the long-term default — warning that could restrict defenders from getting their hands on cutting-edge tools.