OpenAI has taken a significant step in the development of its language models by opening limited access to the new GPT-5.6 family. It includes three models: the flagship Sol, the versatile Terra, and the budget-friendly Luna. Only a small group of trusted partners have gained access via the API and Codex platform. This approach is not accidental but a direct result of a request from US authorities, to whom OpenAI presented its plans and the capabilities of the new models in advance.

The initial phase is structured as a strict preview, and all information about the partners has been shared with the government. The company emphasizes that it does not consider this arrangement a long-term norm and expects to make GPT-5.6 Sol, Terra, and Luna publicly available in the coming weeks. However, at this stage, safety and control are the priorities.

Specifications and Positioning: From Flagship to Budget Solution

Sol is positioned as the most powerful OpenAI model to date. It features two new reasoning modes: max, which allocates more time for deep task processing, and ultra, which uses sub-agents to accelerate complex operations. Terra, on the other hand, is the "workhorse" for everyday tasks. According to OpenAI, it delivers performance comparable to GPT-5.5 but at half the cost. Luna is the most affordable model in the lineup, designed for fast and inexpensive computations.

Performance and Benchmarks: New Records

Sol has already set new standards. On the Terminal-Bench 2.1 test for command-line tasks, it achieved a record result. On GeneBench v1, Sol outperformed GPT-5.5 while using fewer tokens. In the ExploitBench benchmark, the model demonstrated competitiveness with Mythos Preview, consuming only about a third of the output tokens. On ExploitGym tests, all three models improved their performance in cyber tasks as reasoning depth increased.

Safety: The Strongest Security Stack

OpenAI states that GPT-5.6 Sol, Terra, and Luna have received the most powerful security stack in the lineup. Sol does not cross the Cyber Critical threshold within the Preparedness Framework. In tests with Chromium and Firefox browsers, the model found bugs and exploitation primitives but could not autonomously create a fully functional exploit under verified conditions. Over 700,000 GPU-hours in A100 equivalent were used for automated red teaming aimed at finding universal jailbreaks. During the preview phase, multi-layered measures are in place, including model-level restrictions, real-time checks, account-level signals, and monitoring.

Pricing and Future Plans

The new naming scheme is simple: the number denotes the model generation, while Sol, Terra, and Luna represent capability levels. The price per 1 million tokens is $5 input and $30 output for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna. In July, OpenAI plans to launch GPT-5.6 Sol on the Cerebras platform with speeds of up to 750 tokens per second, though access at this stage will also be limited.

My analysis: This move by OpenAI is a clear example of how regulatory pressure is beginning to directly influence the product strategies of major AI companies. Limited access and the sharing of partner data with the government set a precedent that could become the new norm for the entire industry. At the same time, the focus on safety and performance, especially in cyber tasks, indicates that OpenAI is preparing to deploy models in critical infrastructure where the cost of error is extremely high. The market should closely monitor how these models behave in real-world conditions once restrictions are lifted.