OpenAI introduced the GPT-5.6 family in July — Sol, its most powerful model; Terra, balancing efficiency and power; and Luna, built for speed and affordability — with a rollout deliberately limited at the behest of the U.S. government, marking the first time Washington has treated a commercial AI model as a product requiring review before wide release.
The models
GPT-5.6 promises more intelligence per token and stronger performance per dollar, with Sol previewed in late June carrying OpenAI's most advanced safety stack and stronger capabilities in coding, science, and cybersecurity. The three-tier naming — Sol, Terra, Luna — mirrors the industry's convergence on a flagship/workhorse/fast triad, the same structure Anthropic runs with Opus, Sonnet, and Haiku.
The precedent
The government-limited rollout is the story that will outlast the benchmarks. Frontier model releases have been unilateral corporate decisions since the beginning of the modern AI era; GPT-5.6's staged availability, shaped by federal review, establishes a template closer to how pharmaceuticals or export-controlled technology reach market. Notably, the same month saw Washington lift restrictions on Anthropic's most powerful models — evidence the review process runs in both directions.
Also in July
OpenAI shipped GPT-Live-1 and GPT-Live-1 mini on July 8, full-duplex voice models designed for natural real-time conversation with extensive red-teaming behind their safety claims. Voice, long the demo-ware of the AI era, is quietly becoming production infrastructure — and the enterprise contact-center market is the obvious first casualty.