OpenAI unveiled three new models on Friday. They are called Sol, Terra and Luna. Together they form the company’s GPT 5.6 family. They are more powerful than anything the company has released before. And almost nobody can use them yet.
At the request of the United States government, OpenAI agreed to limit the initial release to roughly 20 organisations. These are companies and institutions whose involvement has been formally shared with and approved by the American administration. A wider release to paying customers and the general public is expected within the coming weeks. No exact date has been given. The three models are not identical. They are designed for different purposes and priced differently.
Sol is the most powerful of the three. OpenAI describes it as the strongest model it has ever built. It is built for demanding and complex tasks. These include extended writing and coding sessions, detailed scientific research and work involving computer security. It costs $5 for every million words fed into it and $30 for every million words it produces in response.
Terra sits in the middle. It is designed for everyday business use. OpenAI says it delivers performance similar to its previous top model but at roughly half the cost. That makes it the practical choice for organisations running large volumes of requests every day.
Luna is the fastest and cheapest of the three. It is built for tasks that require speed and volume rather than depth. It costs $1 per million words in and $6 per million words out. It is the entry point in the new family.
The naming approach is deliberate. Sol, Terra and Luna are Latin words for the sun, the earth and the moon. OpenAI said the names are intended to give users a clearer sense of where each model sits in the range. The old naming system, using words like nano and mini, had become confusing as the differences between models grew more meaningful.
Sol also introduces two new ways of working. The first is a maximum reasoning mode that gives the model more time to think through complex problems before producing an answer. The second is called Ultra mode. It splits large tasks across several smaller processes running simultaneously, allowing it to handle work that would overwhelm a single model working alone.
The decision to hold back a broad release came directly from the American government. OpenAI’s chief executive, Sam Altman, confirmed on Friday that the administration had requested a limited preview rather than the company’s originally planned wider launch. He said OpenAI hoped to make the models widely available as quickly as possible.
In a formal statement, OpenAI made its position plain. “We don’t believe this kind of government access process should become the long-term default,” the company wrote. “It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.” It described the restricted release as a short-term step taken while both sides work to build a new framework for handling future releases of this kind.
The political context matters here. President Donald Trump signed an executive order on the 2nd of June requiring government agencies to assess the capabilities of new advanced models before they are released to the public. That review process was given 30 days, meaning a formal framework was due to be in place by the 2nd of July. OpenAI’s Friday launch sits inside that review window. The company says it spent the past month sharing the models’ capabilities with the government, including meetings between Altman and White House officials in early June.
The launch of GPT 5.6 arrived on the same evening that the government separately told Anthropic it could restore limited access to its own advanced model, Mythos 5, which had been shut down two weeks earlier under a different government order. The two announcements arriving within hours of each other underlined how deeply the government has become involved in deciding who can access the most powerful models and when.
Safety was a central theme of OpenAI’s announcement. The company said Sol had been put through more than 700,000 hours of automated testing designed to find weaknesses before release. It said Sol was better at helping users identify and fix problems in computer systems than it was at carrying out attacks. The company said the programme did not meet the definition of what it calls a critical cybersecurity risk, a threshold it defines as bringing genuinely new and dangerous capabilities into existence.
By August, the American government is expected to have established a formal classified process for assessing models of this kind. Until that process exists and is tested, companies releasing powerful new models will face pressure to seek government approval before going public. OpenAI complied this time. It was careful to say that compliance should not be mistaken for agreement.
