OpenAI just dropped its newest model family, GPT-5.6, and the internal reaction is telling. Noam Brown, a senior researcher at the company, said he’d prefer GPT-5.6 over most human research interns for AI tasks.
The limited preview launched on June 26, featuring three variants: Sol, the flagship; Terra, a balanced option splitting the difference between cost and performance; and Luna, the budget-friendly model. You can’t use any of them yet, unless you’re one of approximately 20 vetted partners OpenAI hand-picked for early access.
A government-supervised launch
The staggered rollout wasn’t OpenAI’s first choice. The US government specifically requested the cautious approach, citing concerns around national security and cybersecurity risks. Public availability is scheduled for July 9, 2026, after additional testing to satisfy safety compliance standards.
GPT-5.6 Sol reportedly carries the strongest safety stack OpenAI has ever shipped, following extensive red-teaming conducted over several weeks. The capabilities getting the most attention span coding, biology, and cybersecurity, with each model in the family designed to handle increasingly complex tasks in those domains.
