OpenAI states that, at the request of the United States Government, the company's latest-generation GPT-5.6 model is currently open only to “a few credible partners”. The list of partners has been shared with the Government. At the same time, the company stressed that this was only a short-term arrangement and that there were plans to expand its availability in the coming weeks.
The limited GPT-5.6 series includes the flagship model Sol, the day-to-day use of Terra and the faster and lower cost of Luna. Although Sol is the strongest model of OpenAI's current capacity, this time the Trump government limited the entire product line, not the single model.
The government wants to tighten the front-line model release.
Reports indicate that the United States Government has recently stepped up its intervention on advanced AI system releases. Prior to this, Anthropic, following the roll-out of its high-performance open model Fable 5, had also adjusted the scope of the visit at the request of the Government. According to the report, the request had been related to restrictions on visits by foreign citizens, which had eventually led to Anthropic getting the model offline.
This trend has also led to external discussions on the boundaries of government authority. Dean Ball, who was a White House AI consultant and who was about to join OpenAI, argued that the recent Trump executive order requiring some AI to volunteer for government review no more than 30 days before issuing the most advanced model might in fact make “voluntary review” a de facto compulsory permit.
OpenAI says restrictions should not be the norm.
OpenAI did not hide its dissatisfaction with this arrangement in its statement. According to the company, the current preview limitation is only a “short-term measure” designed to drive GPT-5.6 into a wider distribution phase in the coming weeks.
OpenAI also indicated that the company was discussing with the United States Government a new framework for executive orders, with a focus on network security and duplicated processes for future model release. In their view, restrictive publication should not become the norm, and follow-up would require a clearer mechanism.
New model master code, bio and cyber security capabilities
According to OpenAI, GPT-5.6 Sol has been further enhanced in its proxy capacity for tasks such as code, biological and network security, and has added a more robust mode of reasoning. According to the company, the code workflow of the model performed slightly better than Claude Mythos 5 in part of the baseline test than in Anthropic, while the output token used approximately one third of the latter preview.
In terms of security design, OpenAI states that Sol has adopted the firm ' s strongest security system to date, focusing on enhanced protection against confrontational attacks and favouring more defensive network security than offensive use. The company also claims that these safety constraints are directly embedded in model behaviour itself, rather than relying on additional outer filters.
In terms of availability, GPT-5.6 was initially open to only a few partners, but OpenAI plans to extend the model to ChatGPT, Codex and API users. In terms of pricing, Sol has an input price of US$ 5 per million token and an output price of US$ 30; Terra is about half; and Luna is US$ 1 and US$ 6 respectively. At the same time, the company stated that the reminder cache had been improved to reduce the cost of duplicate requests.