OpenAI has officially launched its GPT-5.6 series of models, comprising the flagship Sol, the balanced Terra, and the lightweight Luna versions.

In a departure from previous public releases, OpenAI has not made the models immediately available to all. Instead, acting upon a request from the U.S. government, it is initially providing preview access only to a select group of "trusted partners."

OpenAI stated this move is to assist the U.S. government in establishing a safety evaluation process for cutting-edge AI models. However, the company also took the unusual step of publicly stressing that a government approval-based model release mechanism "should not become the long-term default." This statement signals that, against the backdrop of the Trump administration's increasing focus on national security oversight for advanced AI, the release of models by U.S. AI companies has entered a new phase. The focus is not only on model capabilities as a competitive edge but also on the models themselves undergoing government safety reviews.

Introducing the GPT-5.6 Series: Sol, Terra, and Luna

On Friday, 26th Eastern Time, OpenAI unveiled the three models in the GPT-5.6 series. These are described as "our next-generation frontier model, GPT-5.6 Sol; a balanced model for efficiency and everyday work, GPT-5.6 Terra; and a fast, cost-effective model for high-frequency, large-scale tasks, GPT-5.6 Luna."

Among them, the Sol model introduces new, more advanced capabilities. OpenAI indicated the new models show significant improvements in areas such as software engineering, autonomous agent tasks, cybersecurity, and defensive research. They have undergone extensive safety testing, including over 700,000 GPU-hours of automated testing and substantial external "red team" evaluations.

Regarding pricing, GPT-5.6 Sol is priced at $5 per million input tokens and $30 per million output tokens. This cost is approximately half that of Anthropic's Claude Fable 5 model, which is priced at $10 and $50 per million tokens for input and output, respectively. Terra is priced at half the cost of Sol, and Luna is priced at less than half the cost of Terra.

The Most Significant Change: A Government-Mandated Limited Release

More than the model performance, the release strategy has drawn significant market attention.

OpenAI announced that the GPT-5.6 series is currently only available for preview to a limited number of trusted partners. This arrangement was not OpenAI's own initiative but was implemented in response to a request from the Trump administration. OpenAI disclosed this in its announcement.

Reports indicate the U.S. government aims to establish a unified review process for frontier models with potential cybersecurity capabilities before they are fully public. Consequently, it requested OpenAI adopt a phased release strategy, initially allowing only approved organizations to access the models. The first batch of companies granted access numbers around twenty, with the scope expected to widen gradually over the coming weeks.

Under the current mechanism, it is reported that access for some clients even requires individual government approval.

This approach aligns with the Trump administration's recent stricter regulatory stance on frontier AI models. Previously, the U.S. government had requested Anthropic to retract the public release of some new models and established an evaluation framework for "Covered Frontier Models."

OpenAI Voices Concerns Over Government Approval, Advocates for a Repeatable Process

While complying with the U.S. government's request for a limited release, OpenAI included a notably strong statement in its official announcement.

The announcement first clearly stated that OpenAI "firmly believes in enabling broad access and plans to fully launch the GPT-5.6 Sol, Terra, and Luna models in the coming weeks." It noted that during the preview period, testing and close collaboration with partners would continue to advance broader availability. The announcement then added a key passage.

This statement is one of the most closely watched aspects of OpenAI's announcement. The company explicitly stated its support for the Trump administration's goal of safeguarding national security. However, it also emphasized that frontier AI models should be made available to developers, research institutions, businesses, and cybersecurity defenders as soon as possible, as these groups need timely access to advanced AI capabilities.

OpenAI further expressed its hope that, through this preview, it can work with the U.S. government to establish a clear, repeatable approval process, rather than having government case-by-case approvals become the long-term operating norm for the future AI industry.

Model Safety Assessment: Not Yet at the Highest Risk Level

OpenAI also released the latest results from its deployment safety assessment.

According to its Deployment Safety report, while GPT-5.6 shows enhanced capabilities in autonomous task execution, vulnerability analysis, and complex reasoning, the company believes it has not yet reached the dangerous threshold in its Preparedness Framework that would necessitate the highest level of restriction.

OpenAI stated: the model's cybersecurity capabilities are significantly enhanced; it is better at discovering software vulnerabilities than at aiding in attacks; it has not reached a danger level requiring full deployment restrictions; and it underwent automated testing, expert red teaming, and external agency evaluations prior to release.

Codex and ChatGPT to Also Upgrade to GPT-5.6

Beyond the API, OpenAI indicated that its Codex and ChatGPT products will also be upgraded to the GPT-5.6 architecture.

This upgrade signifies that OpenAI's entire product ecosystem—including ChatGPT, Codex, and the API—will gradually transition to the GPT-5.6 architecture. However, the timeline for a full public rollout remains dependent on the progress of the ongoing government review process.

AI Competition Enters a New Phase: Beyond Models, Navigating Regulation

Industry observers believe the most significant implication of this GPT-5.6 release may not be another leap in model performance, but rather the U.S. government's first deep involvement in the release process of the world's most advanced AI models.

Following Anthropic, OpenAI now also requires government safety evaluation before launching its models, indicating that U.S. AI regulation is gradually shifting from traditional industry oversight to managing the deployment of frontier models.

However, OpenAI's public emphasis that it does not believe a government approval process should become the long-term default also shows the company hopes the current arrangement is merely a transitional mechanism for a special period, not a permanent rule for the future AI industry.

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