OpenAI has confirmed it is delaying the public release of its next-generation AI models, GPT-5.6, at the direct request of the Trump White House. The company will first share the models with a small, government-preapproved set of customers, then work with the administration to gradually expand access over the coming weeks.

OpenAI is privately unhappy with the arrangement, according to a person familiar with the company's thinking, but views it as a short-term concession rather than a lasting precedent.

"We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." — OpenAI blog post

A Framework That Doesn't Exist Yet

Earlier this month, President Trump signed an executive order designed to address cybersecurity concerns around powerful new AI models. The order proposed a voluntary 30-day advance-sharing process between AI labs and the government — with an explicit carve-out against turning that process into a de facto licensing regime.

The problem: that voluntary framework hasn't actually been built yet. OpenAI executives confirmed on Friday that no formal process is in place, leaving frontier AI labs in a regulatory grey zone where cooperation with the government is technically voluntary but practically unavoidable.

Anthropic Still Caught in Its Own Standoff

OpenAI's delay comes just two weeks after the White House sent an export control directive to Anthropic, which forced the company to take its most advanced models offline for all customers. That dispute remains unresolved — some of Anthropic's own employees are still blocked from accessing its most capable AI systems.

The back-to-back interventions are creating significant uncertainty across the US AI industry, particularly given that the Trump administration has spent much of the past two years cutting AI regulation to boost American competitiveness against China.

What GPT-5.6 Actually Offers

OpenAI's GPT-5.6 will launch in three variants:

  • Sol — the most capable version, with top benchmark scores in cybersecurity, biology, and agentic tasks
  • Terra — a mid-tier option
  • Luna — a fast, cost-efficient version

Alongside the new models, OpenAI says it has developed a "layered safeguard stack" aimed at preventing misuse — including cyberattacks and other malicious applications.

Opaque Approval Process

OpenAI plans to expand its approved customer list as early as next week, including some international partners. However, the company says it cannot disclose how the White House is making approval decisions. The process, as described by OpenAI executives, is straightforward but opaque: the company submits a list of customers and receives feedback from the government.

The White House did not respond to requests for comment.