The GPT 5.6 limited release will look nothing like a typical OpenAI product launch. Rather than distributing the model broadly, OpenAI plans to share it only with a select group of close partners during a preview period, after the Trump administration asked the company to hold back, according to reporting by The Information.
At a staff meeting this week, chief executive Sam Altman told employees that the government would be ‘approving access customer by customer’ during the preview window. If that phase goes smoothly, Altman reportedly said OpenAI hopes to follow with a general release ‘a couple of weeks later.’
The GPT 5.6 Limited Release and What It Signals
The agencies behind the request were the Office of the National Cyber Director and the Office of Science and Technology Policy. OpenAI’s staffers reportedly ‘worked closely’ with the government on the upcoming release, suggesting this was less a directive and more a negotiated arrangement.
That arrangement is now formalised in law. On 2 June 2026, President Trump signed an executive order titled Promoting Advanced Artificial Intelligence Innovation and Security, which establishes a voluntary framework permitting developers of advanced AI models to provide the federal government with model access prior to public release, with a review window of up to 30 days.
That final figure was itself a compromise. According to the New York Times, an earlier draft of the order would have created a review window of up to 90 days. Trump pulled back from the longer window just hours before signing.
The companies in the room for these negotiations were named in a PYMNTS report on a White House Office of the National Cyber Director briefing that preceded the executive order. OpenAI, Anthropic, and Reflection AI attended alongside cloud providers, semiconductor firms, cybersecurity companies, and banks. During that process, AI companies lobbied to share their models with the government two weeks in advance of any release.
This is a striking turn for an administration that came to power positioning itself as taking a ‘hands-off’ approach to AI regulation. The executive order also directs the Treasury Department to establish an AI cybersecurity clearinghouse, and instructs the Secretary of Homeland Security, through the CISA director, to issue Binding Operational Directives covering the cyber defence of civilian federal government information systems.
Anthropic’s Glasswing Model Is the Template Everyone Is Watching
The GPT 5.6 limited release puts OpenAI on a path Anthropic has already been walking. Earlier this year, Anthropic announced that its new frontier cyber model, Claude Mythos, would be released only to a small group of partners through a programme called Project Glasswing. The company argued that the model was too powerful to release broadly without controls.
The results of that restricted approach are now quantifiable. Anthropic’s newsroom reports that Project Glasswing partners have found more than 10,000 high- or critical-severity security flaws so far. Alongside Mythos, Anthropic has also released Claude Security, a product built on its latest public frontier models including Claude Opus 4.8, to scan codebases and suggest patches.
The programme is expanding. Cybersecurity Dive reports that Anthropic has extended Project Glasswing to 150 more organisations, including critical infrastructure operators. Whether that constitutes prudent stewardship or a form of controlled market-building is a debate that has not resolved.
The underlying concern with frontier cyber tools is concrete regardless of the commercial angle. Large language models have proven capable of writing malware, and some can execute ransomware attacks autonomously. Models like Mythos are ostensibly able to identify and exploit software vulnerabilities at speeds that outpace any human analyst. Since enterprise networks frequently contain undiscovered bugs that serve as entry points, the asymmetry between attacker speed and defender awareness is the crux of the problem.
Reuters reported that OpenAI, Anthropic, and Alphabet’s Google all met with the US government about cybersecurity as the executive order was being developed, according to a senior official who spoke in May. That breadth of engagement suggests the White House’s interest extends well beyond one company’s launch calendar.
The more immediate question is whether the 30-day review window is long enough to catch meaningful risks in a model as complex as GPT 5.6. If the limited release proceeds without incident, OpenAI’s broader rollout could follow within weeks. If the review surfaces concerns, the framework for extending that window, or tightening the terms entirely, is now in place.
