OpenAI Begins Limited Preview of GPT-5.6 for Trusted Partners

OpenAI has begun a limited preview of its new GPT-5.6 model family, giving a small group of trusted partners early access ahead of a broader public release expected in the coming weeks.

The new lineup includes three variants - Sol, Terra, and Luna - each designed for different use cases and pricing tiers. According to OpenAI, Sol is its most capable AI model to date, while Terra is aimed at everyday applications with performance comparable to GPT-5.5 at roughly half the cost. Luna is positioned as the company's most affordable model for developers seeking lower operating costs.

The preview comes after OpenAI shared GPT-5.6 with the U.S. government under a voluntary review process established for advanced AI models.

Three Models for Different AI Workloads

Terra is designed to balance performance and cost, targeting mainstream AI workloads, while Luna offers the lowest pricing in the GPT-5.6 family for applications that prioritize affordability over maximum capability. OpenAI plans to make all three models more broadly available in the coming weeks, although it has not announced an exact release date.

Government Review and Stronger AI Safeguards

Before launching the preview, OpenAI provided U.S. government officials with access to GPT-5.6 as part of a voluntary evaluation process. According to the company, the limited release to trusted partners was made at the request of the U.S. administration so the model could complete government review before a wider rollout.

OpenAI said it does not believe this level of government access should become the long-term standard but described the current approach as a temporary step to help accelerate public availability.

The review follows an executive order signed by U.S. President Donald Trump earlier this month encouraging developers of the most advanced AI models to voluntarily submit them for government cybersecurity evaluation approximately 30 days before public release. According to a recent report by The New York Times, OpenAI, Anthropic, Google, xAI, and Microsoft had already been providing early access to their latest AI models before the executive order was issued.

Focus on Cybersecurity

OpenAI said GPT-5.6 includes stronger safeguards against misuse, particularly in cybersecurity-related scenarios.

The company said it spent several weeks testing Sol against real-world attack techniques before launch and added additional protections for high-risk requests. Across the GPT-5.6 family, OpenAI has also trained the models to refuse prohibited cybersecurity assistance, including attempts to bypass built-in safety restrictions through jailbreaks.

According to the company, approximately 700,000 GPU hours were used to identify universal jailbreak techniques and develop countermeasures. OpenAI also said it has established a rapid-response process to investigate and address newly discovered jailbreak methods.

The company's emphasis on AI safety follows recent scrutiny of advanced AI models. Earlier this month, Anthropic temporarily suspended access to its Mythos and Fable models after U.S. government concerns over potential misuse. Access has since been partially restored for a limited group of organizations.

OpenAI expects to expand access to GPT-5.6 in the coming weeks, positioning the new model family as its latest effort to balance higher AI performance with stronger security safeguards and more flexible pricing for developers.