OpenAI Launches GPT 5.6 Models Featuring Sol Terra and Luna Naming Structures with Advanced Reasoning and Government Regulatory Coordination
OpenAI just launched its 6 th iteration of models recently, a mere 2 months after the launch of GPT 5.5. The company has then unexpectedly unveiled its new naming structure for the technology and the capability groups it packages into. The number indicates the total generation of the neural network, but the company will also be listing a number of different names for various capability groups, with Sol as the flagship, Terra a medium optimal choice for application used daily, the Luna as a high speed, low cost utility.
And the cost of Operatios for the middle tier Terra solution is highly competitive when compared to the old GPT 5.5 flagship, but at 2x the operaing cost. For access for Developers, OpenAI has set the prices per 1M tokens of input and output data; for example, Sol comes at $5 and $30, respectively, for input and output data; Terra at 250$ and 1500$, respectively; the lighter Luna at just 1$ to 6$ for output. Moreover, OpenAI has teamed up with Cerebras to launching Sol on July, and stated they plan to achieve processing speeds of up to 750 tokens per second for some enterprise users.
In order to handle extremely advanced processing tasks flagship Sol offers a Max Reasoning mode, giving it many times more time to compute a serious problem before responding; Additionally Ultra mode has been added, which runs a web of multi agent subagents that collaborate on multi step processes.
These improvements have produced quantifiable victories in a variety of fields. Sol set the same high bar as the best past software engineering models on the Terminal Bench 2.1 benchmark, which tests command line navigation and tool integration. Sol beat out GPT 5.5 on the GeneBench v1 benchmark in genomics and quantitative biology while using fewer output tokens and handled backtesting calculations faster than any other GPD model used in cybersecurity experiments on ExploitBench while needing only approximately ⅓rd the amount of output tokens per system exploit fix.
In this generation, safety and defense were the first priorities. OpenAI used over 700,000 A100 equivalent GPU hours for red teaming it automatically to find possible weaknesses of the system and to prevent jailbreaks. The security stack is layered.
When it generates the output, a real time classifiers can block wrong outputs if it determines that the output contains a possible policy violation.
At that point, the generation halts while a larger reasoning model considers the context of the conversation before producing a result to present to the user.
The launch of the GPT 5.6 series is significant for its apparent close association with the US government. Following US Government request, OpenAI are releasing a very limited preview to a small number of trusted partners. OpenAI articulated its position in the official documentation on the phased roll out as follows
At their request, we are starting with a limited preview for a small group of trusted partners whose participation has been shared with the government, before releasing more broadly. During this preview, we will continue testing and coordinating closely with partners as we work toward broader availability. We do not believe this kind of government access process should become the long term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.
Although the first preview will be limited, OpenAI intends to open the gates to more developers over the next few weeks via its API, Codex, and ChatGPT platforms while it enhances its defensive mechanisms with learnings from this early testing.
