OpenAI has taken a significant step in the development of its technologies by opening limited early access to the new GPT-5.6 model family. It includes three versions: the flagship Sol, the versatile Terra, and the budget-friendly Luna. Trusted partners were the first to gain access via API and Codex.

Why limited access?

The company decided to launch the preview in coordination with U.S. authorities, presenting the plans and capabilities of the models to the government in advance. Information about the partners was shared with regulators. OpenAI emphasizes that this format will not become a long-term norm, but at this stage, it is a necessary step to ensure safety.

Key differences between the models

Sol is positioned as the most powerful model in the lineup to date. It features new reasoning modes: max, which allows more time for deep analysis, and ultra, which uses sub-agents to accelerate complex tasks. Terra is a "workhorse" for everyday tasks, which, according to OpenAI estimates, delivers performance on par with GPT-5.5 but costs half as much. Luna is the fastest and most affordable option, designed for lightweight use cases.

Technical achievements and safety

Sol set a new record in the Terminal-Bench 2.1 benchmark for command-line tasks. On GeneBench v1, it outperformed GPT-5.5 while using fewer tokens, and on ExploitBench, it achieved results comparable to Mythos Preview using only a third of the output tokens. All three models showed improved performance in cyber tasks on ExploitGym tests as reasoning depth increased.

Special attention has been paid to safety. GPT-5.6 Sol does not cross the Cyber Critical threshold within the Preparedness Framework. In tests with Chromium and Firefox, the model found bugs but could not autonomously create a fully functional exploit. Over 700,000 A100-equivalent GPU hours were used for automated red-teaming to verify defenses. During the preview phase, multi-layered measures are in place, including model-level restrictions, real-time checks, and monitoring.

Pricing and future plans

The cost per 1 million tokens is: for Sol — $5 for input and $30 for output, for Terra — $2.50 and $15, for Luna — $1 and $6. 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 comment: OpenAI is clearly diversifying its offerings, tailoring models to different budgets and tasks. Sol looks like a serious step forward in reasoning and cybersecurity, but the limited access and close collaboration with regulators suggest the company is preparing for stricter AI oversight. I expect that after the public launch, these models will become a standard for developers, especially in the automation and security niche.