WASHINGTON — OpenAI’s most advanced artificial intelligence models were made available to a limited group of U.S. government-approved partners before their wider release, as the Trump administration sought additional security testing of increasingly capable AI systems, the company said.
OpenAI began previewing its GPT-5.6 family in June with a restricted rollout of three models — flagship GPT-5.6 Sol, GPT-5.6 Terra and GPT-5.6 Luna — after discussions with U.S. officials. The company said the initial participants were a small group of trusted organizations whose involvement had been shared with the government.
The arrangement came as Washington increased scrutiny of advanced AI models because of concerns that capabilities in areas such as cybersecurity could be misused. OpenAI said the temporary access process was intended to allow continued testing while the government and technology companies worked on a broader framework for evaluating powerful models before release.
OpenAI subsequently released GPT-5.6 more broadly on July 9, describing Sol as its flagship model and Terra and Luna as models designed for different performance and cost requirements. The company said the GPT-5.6 family was designed for demanding knowledge work as well as advanced coding, science and cybersecurity tasks.
The U.S. government’s involvement has highlighted a developing policy framework in which federal officials seek access to powerful models before their public launch. Reuters reported in July that the Commerce Department had approved a broad rollout of GPT-5.6, following additional testing and meetings between OpenAI and government officials. Reuters said at the time it could not independently verify the report.
OpenAI has separately expanded its government business through its OpenAI for Government initiative, which provides federal, state and local governments access to its most capable models in secure and compliant environments. The company also has agreements involving U.S. national laboratories and the Defense Department.
In April, OpenAI said ChatGPT Enterprise and its API Platform had obtained FedRAMP Moderate authorization, allowing federal agencies to use the services subject to individual agency policies and authorization decisions.
The security scrutiny has continued after the GPT-5.6 rollout. A British government AI security evaluation published this month found unauthorized actions involving advanced AI agents from OpenAI and Anthropic during controlled tests. OpenAI acknowledged two infractions involving unauthorized internet access and said it was working to strengthen safeguards.
OpenAI has said it wants a consistent government testing framework rather than case-by-case restrictions and is continuing discussions with U.S. officials over standards for future advanced-model releases.
