OpenAI Unveils GPT-5.6 Sol, Terra and Luna With Access Limited by US Government Review

OpenAI has introduced its next-generation GPT-5.6 family of artificial intelligence models, but the company’s most advanced systems will remain unavailable to most users following a request from the U.S. government to delay their broad public release.

The new GPT-5.6 lineup consists of three models: Sol, the flagship and most capable version; Terra, a balanced model designed for general-purpose use; and Luna, a faster and more affordable option. OpenAI said the restrictions are temporary and that it is working with the Trump administration to gradually expand access to a wider group of customers, including international partners, beginning next week.

GPT-5.6 Sol Leads the New Model Family

OpenAI has named the three models after celestial bodies, with Sol representing the Sun, Terra referring to Earth and Luna representing the Moon.

Among the new releases, GPT-5.6 Sol is the company’s highest-performing model. According to OpenAI, it delivers improved capabilities in coding, biology, cybersecurity and agentic tasks. The model includes two operating modes: a “max” reasoning mode and an “ultra” mode that coordinates multiple sub-agents to tackle highly complex problems.

OpenAI said benchmark testing showed GPT-5.6 Sol performed strongly across cybersecurity, biology and coding evaluations. The company also stated that the model slightly outperformed Anthropic’s Claude Mythos 5 on coding workflows while achieving comparable performance using only one-third of the output tokens.

Public Release Delayed by Government Request

OpenAI confirmed that the public launch of GPT-5.6 has been postponed indefinitely following a request from the U.S. government. At present, only a limited number of customers that have received government approval will be able to access the models.

The company said it is working with the Trump administration to establish a process that would allow broader availability in the coming weeks. However, the criteria used by the government to approve customers have not been publicly explained. According to reports, OpenAI submits a list of prospective customers and receives feedback from government officials.

OpenAI said it does not believe this type of government access process should become the long-term standard. The company argued that restricting access prevents developers, enterprises, cybersecurity professionals and international partners from benefiting from its latest technology. Nevertheless, it said it views the current arrangement as the best path towards wider availability while discussions continue on a future framework for frontier AI releases.

Chief Executive Officer Sam Altman also described the process as “not optimal” but said OpenAI wanted to remain a dependable partner while continuing to pursue its mission of benefiting humanity.

Safety Features and Pricing

OpenAI said GPT-5.6 Sol incorporates its most comprehensive security framework to date. The company explained that the model has been trained to resist jailbreaking attempts and adversarial attacks while prioritising defensive cybersecurity applications over offensive uses.

Unlike approaches that rely on external filtering systems, OpenAI said the model’s safety guardrails have been integrated directly into its core behaviour.

The company also announced that, once generally available, GPT-5.6 models will be accessible through ChatGPT, Codex and the OpenAI API for paid users. Sol will cost $5 per million input tokens and $30 per million output tokens. Terra will be available at half that price, while Luna will cost $1 per million input tokens and $6 per million output tokens. OpenAI added that improvements to prompt caching will make repeated prompts more efficient and predictable.

Executive Order Shapes AI Model Releases

Earlier this month, U.S. President Donald Trump signed an executive order aimed at addressing the cybersecurity risks associated with advanced frontier AI models. The order proposes a voluntary framework under which AI developers can share models with the U.S. government 30 days before public release.

The executive order states that the review process is voluntary rather than a licensing system. OpenAI and Anthropic had already been working with the U.S. government to evaluate unreleased AI models, while Google DeepMind, Microsoft and xAI agreed to provide early access to future models for national security assessments.

Meta remains the only major U.S.-based AI model developer that has not agreed to voluntarily share its models with the government for review, according to the report.

Critics of the expanding government oversight argue that the absence of clearly defined safety standards could result in prolonged launch delays, potentially affecting investment in AI infrastructure and influencing global competition in artificial intelligence.

With inputs from Reuters