OpenAI released the GPT-5.6 series of models, with limited access as requested by the U.S. government, stating outright that this should not become the long-term default approach.

OpenAI officially releases the GPT-5.6 series of models, including the flagship Sol, the balanced Terra, and the lightweight Luna. Unlike previous public releases, OpenAI did not immediately open them fully this time; instead, at the request of the U.S. government, it only provides preview access to a small number of “trusted partners.”

OpenAI said this is to support the U.S. government’s establishment of a safety assessment process for cutting-edge AI models. At the same time, in a rare public emphasis, it stated that the government-approval-based model access mechanism “should not become a long-term default mode.” This signals that, amid the Trump administration’s ongoing strengthening of national-security oversight for cutting-edge AI, U.S. AI companies’ model releases have entered a new phase: not only model capability has become the competitive focus, but the model release process itself has also started to face government security reviews.

Positioning of the GPT-5.6 Sol, Terra, and Luna Models

On Friday the 26th, Eastern Time, OpenAI simultaneously announced three models in the GPT-5.6 series, including “our next-generation frontier model GPT-5.6 Sol; the balance model GPT-5.6 Terra that weighs efficiency alongside day-to-day work; and a fast, economical model GPT-5.6 Luna for high-frequency, large-scale tasks.”

• GPT-5.6 Sol: Positioned as the flagship model, for the most complex reasoning, scientific research, software development, cybersecurity, biological research, and Agent workflows.
• GPT-5.6 Terra: Focuses on comprehensive capabilities; compared with GPT-5.5 it maintains performance that is close, but with costs reduced by about 50%, targeting enterprise and daily productivity scenarios.
• GPT-5.6 Luna: The fastest and lowest priced, suitable for large-scale online services and high-throughput scenarios.

Among them, Sol adds higher-tier modes, including:

• Max: Longer time to think, deeper reasoning;
• Ultra: Able to coordinate multiple sub-Agents to complete complex long-term tasks.

OpenAI said the new models have achieved significant improvements across software engineering, autonomous agent (Agent) tasks, cybersecurity, and defensive research, and have undergone more than 700,000 GPU hours of automated security testing plus extensive external red-team evaluations.

In terms of service pricing, GPT-5.6 Sol is charged at $5 per million input tokens and $30 per million output tokens. This cost is only about half of Anthropic’s Claude Fable 5 model, which charges $10 per million input tokens and $50 per million output tokens. Terra is priced at half of Sol, and Luna is priced at less than half of Terra.

The biggest change is the release method: limited access required by the U.S. government

More than model performance, what has attracted greater market attention is this release method.

OpenAI announced that the GPT-5.6 series is currently only available in preview to a limited number of trusted partners. This arrangement was not decided by OpenAI on its own; it was implemented at the request of the Trump administration. In its announcement, OpenAI disclosed:

“As part of ongoing communication with the U.S. government, before today’s release, we first demonstrated our plans and model capabilities. Per the government’s request, we will provide a limited preview initially to a small group of trusted partners. The participation status of these partners has been reported to the government, and then we will proceed with a broader release.”

The media said the U.S. government wants, before models are fully made public, to establish a unified review process for frontier models with potential cyber capabilities, so it requires OpenAI to adopt a phased release strategy—allowing only approved institutions to access the model first. Approximately twenty companies were among the first to receive access, and the scope of openings will be expanded gradually over the coming weeks.

The reports said that under the current mechanism, some customers’ model access permissions even require government approval one by one.

This arrangement continues the recent stricter regulatory approach by the Trump administration toward cutting-edge AI models. Previously, the U.S. government required Anthropic to withdraw part of a new model’s public release and set up an evaluation system for “Covered Frontier Models.”

OpenAI publicly expresses concerns about government approval, hoping to establish a repeatable release process

Although it complied with the U.S. government’s requirements to implement limited access, OpenAI specifically added a rather firm statement in its official announcement.

The announcement first makes clear that OpenAI “firmly believes in the need for broad access and plans to fully roll out the GPT-5.6 Sol, Terra, and Luna models in the coming weeks,” and says that during the model preview period it will continue testing and work closely with partners to advance broader openness. It then continues:

**“**We believe that this government-involved access process should not become the long-term default mode. It prevents users, developers, enterprises, cybersecurity defenders, and global partners who truly need these top-tier tools from obtaining these resources.

We are taking this short-term measure because we believe it is the best way to achieve a broader release within the coming weeks. At the same time, we will work with the government to formulate an administrative order framework on cybersecurity and establish a set of processes that can be executed repeatedly and applied to future model releases.”**

This is one of the most attention-grabbing statements in OpenAI’s announcement.

On the one hand, the company clearly states support for the Trump administration’s goal of safeguarding national security. On the other hand, it also emphasizes that frontier AI models should be opened to developers, research institutions, enterprises, and cybersecurity defenders as soon as possible, because these groups need timely access to advanced AI capabilities.

OpenAI further said it hopes—together with the U.S. government—to establish a clear, repeatable approval process through this preview, rather than allowing government approval to become a case-by-case, long-term operating approach for the future AI industry.

Model safety assessment: not yet at the highest risk tier

OpenAI also released its latest deployment safety assessment results.

According to the Deployment Safety report, while GPT-5.6 has further improved capabilities in autonomous task execution, vulnerability analysis, and complex reasoning, the company believes it has not yet reached the dangerous thresholds in the Preparedness Framework that require the highest level of restrictions.

OpenAI said: The model’s cybersecurity capabilities have clearly increased; it is better at finding software vulnerabilities rather than helping to carry out attacks; it has not reached the dangerous level that would require fully restricting deployments; and it has undergone automated testing, expert red-team evaluations, and assessments by external institutions before release.

Codex and ChatGPT will also be upgraded to GPT-5.6

In addition to the API, OpenAI said:

• ChatGPT will gradually integrate the GPT-5.6 series;
• Codex will also be upgraded to GPT-5.6 models;
• The developer platform will later open up more access.

Media reports said this upgrade means that OpenAI’s entire future product ecosystem—including ChatGPT, Codex, and the API—will gradually shift to the GPT-5.6 architecture, but the timing of a full opening still depends on how this government review process progresses.

AI competition enters a new phase: not just models, but regulation

Industry insiders believe the greatest significance of this GPT-5.6 release may not be that model performance has improved again, but rather that the U.S. government is, for the first time, deeply intervening in the release process of the world’s most advanced AI models.

After Anthropic, OpenAI also needs to undergo government safety evaluations before model launch, meaning U.S. AI regulation is shifting from traditional industry oversight toward managing the deployment of frontier models.

However, OpenAI’s public emphasis that it does not believe the government-approval process should become a long-term default practice also shows that the company expects the current arrangement to be a transitional mechanism for a special period—not a long-term rule governing how the future AI industry operates.

In the coming weeks, as GPT-5.6 gradually expands access and the U.S. government formally establishes a review system for frontier models, how the AI industry finds a balance between national security and technological openness will become a new focal point of global AI competition.

Risk Warning and Disclaimer

        There are market risks, and investment requires caution. This article does not constitute personal investment advice, and it does not take into account the specific investment objectives, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article fit their specific circumstances. Invest at your own risk based on this.