OpenAI Begins Limited GPT-5.6 Preview After US Government Review Request

OpenAI has started rolling out its latest GPT-5.6 AI models to a limited group of trusted partners, saying the restricted preview follows a request from the U.S. government to review advanced AI systems before their wider release.

In a blog post published on Friday, the company said the preview currently includes only partners "whose participation has been shared with the government." OpenAI described the arrangement as a temporary measure and stressed that it does not believe this type of government access should become the long-term approach for releasing frontier AI models.

The GPT-5.6 family consists of three models: Sol, the company's flagship model; Terra, designed for general-purpose workloads at a lower cost; and Luna, an entry-level model focused on affordability and speed. OpenAI said broader availability through ChatGPT, Codex, and its API is expected in the coming weeks.

OpenAI Questions Long-Term Government Oversight

The limited rollout comes as the U.S. government increases its scrutiny of advanced artificial intelligence systems.

OpenAI said it complied with the request to help accelerate GPT-5.6's broader deployment but argued that prolonged government review could slow innovation and delay access for developers, businesses, cybersecurity professionals, and other users.

According to the company, it is working with the administration to establish a more predictable framework for future AI model releases rather than relying on case-by-case arrangements.

The debate over government oversight has intensified following recent reports that Anthropic temporarily restricted access to some of its latest AI models after concerns were raised about potential misuse. Those developments have sparked wider discussions within the AI industry about how governments should balance national security with technological innovation.

GPT-5.6 Focuses on Performance and AI Safety

OpenAI said GPT-5.6 Sol is its most capable model to date, introducing a new "max" reasoning mode that allocates additional computing time to solve more complex tasks. The company also introduced an "ultra" reasoning mode that uses coordinated AI subagents for particularly demanding problems.

According to OpenAI, Sol delivers improved performance in software development, cybersecurity, and scientific applications while incorporating stronger protections against misuse.

The company said it invested heavily in strengthening GPT-5.6's security before release. Rather than relying solely on external safety filters, OpenAI said many of the safeguards have been integrated directly into the model's behavior. The system has also been trained to refuse prohibited cybersecurity requests, including attempts to bypass safety protections through jailbreak techniques.

OpenAI said it used approximately 700,000 GPU hours to identify vulnerabilities and improve resistance to adversarial attacks. It also pledged to continue monitoring newly discovered jailbreak methods through a rapid-response process.

The limited preview marks another step in OpenAI's strategy to balance increasingly powerful AI capabilities with stronger security safeguards while navigating growing government oversight of advanced AI technologies.