A Three-Model Lineup Built for Different Scales of Work
OpenAI has officially pulled back the curtain on GPT-5.6, its most advanced AI model family to date. The lineup is made up of three distinct models, each positioned for a different tier of use. Sol is the flagship, engineered for the most demanding and complex workloads. Terra sits in the middle, designed for balanced reasoning and everyday tasks. Luna rounds out the family as the fastest and most affordable option, built for high-volume applications where speed and cost efficiency matter most.
Across all three, GPT-5.6 delivers meaningful improvements in areas including coding, scientific reasoning, cybersecurity, biology, and long-running autonomous tasks. Sol also introduces two advanced operating modes not found in the others. Max pushes the model into deeper, more intensive reasoning territory, while Ultra is built specifically to orchestrate sub-agents across sprawling, multi-step workflows.
The Catch: Almost No One Can Use It Yet
The bigger story isn't what GPT-5.6 can do — it's who's actually allowed to try it. OpenAI is rolling out access in a tightly controlled preview, limited to a small group of customers approved by the Trump administration. The models are currently undergoing additional national security reviews as part of a new federal oversight framework, and that process is dictating who gets in first.
OpenAI stated that it previewed its plans and the models' capabilities with the U.S. government ahead of the launch and is starting with a limited release for trusted partners whose participation has been shared with the government. The company frames this as a temporary measure and says it expects to make GPT-5.6 broadly available in the coming weeks, though no specific date has been announced.
Notably, OpenAI was also clear that it does not view this kind of mandatory government pre-approval as a model it wants to adopt long-term for releasing frontier AI.
Safety Infrastructure Behind the Launch
OpenAI paired the GPT-5.6 release with what it describes as its most robust safety stack to date. The deployment includes strengthened real-time protections against high-risk cyber activity and patterns of repeated misuse. Before any of it went live, the model was put through extensive human red-teaming alongside more than 700,000 A100 GPU-equivalent hours of automated safety testing — a significant investment in hardening the system prior to release.
A Pattern Taking Shape Across the Industry
GPT-5.6's restricted rollout isn't happening in isolation. Just weeks earlier, U.S. authorities required Anthropic to restrict access to its Claude Mythos 5 and Fable 5 frontier models, citing national security concerns. Mythos has since returned for a limited group of users, but Fable 5 remains off-limits to the general public, currently available only to approved U.S.-based entities. OpenAI is now navigating a nearly identical situation.
There's an additional layer of pressure behind the cautious approach. Anthropic recently alleged that Chinese tech giant Alibaba used thousands of user accounts to systematically access Claude and distill its outputs to improve Alibaba's own Qwen model family. It's not an isolated claim — similar allegations have surfaced before, and they reflect a growing concern that frontier AI models can be reverse-engineered or exploited before developers have the chance to properly secure them.
Whether or not that concern directly shaped OpenAI's rollout strategy, the direction of travel is clear. Getting the world's most capable AI models to market is no longer a purely technical challenge. It has become a geopolitical one.