The Trump administration has asked OpenAI to limit GPT-5.6’s launch to a select group of government-approved partners and CEO Sam Altman has agreed. Here is what is driving the decision, what it changes, and what it means.
- What Happened: OpenAI Agrees to a Controlled GPT-5.6 Launch
OpenAI’s newest flagship model, GPT-5.6, will not launch the way its predecessors did. Rather than rolling out publicly to all users, the company plans to share it only with a select group of close partners and that decision came directly from the Trump administration.
According to reporting by The Information, confirmed by CNN and Axios, the White House asked OpenAI to limit the release because of GPT-5.6’s advanced cybersecurity capabilities. At an internal meeting this week, CEO Sam Altman told staff that the government would be approving access “customer by customer” during a preview period. If the controlled rollout goes well, OpenAI hopes to follow with a general public release within a couple of weeks.
Who asked? The specific agencies were the Office of the National Cyber Director (ONCD) and the Office of Science and Technology Policy (OSTP) — two of the administration’s key players in AI security, per Axios.
In Altman’s memo to employees, he reportedly stated: the administration made a request, and OpenAI agreed to it — even while making clear this is not their preferred long-term model. The company said it would work with the government and industry to find a more sustainable approach for future releases.
- Why the White House Is Concerned About GPT-5.6
GPT-5.6 is not just another incremental update. Both OpenAI and the administration view the model as on par with Anthropic’s Mythos 5 — a model so powerful the US government already forced it offline on June 12, 2026, using an emergency export control directive.
The core concern is cybersecurity capability. Advanced frontier models at this tier can dramatically accelerate vulnerability discovery, software exploit generation, and adversarial cyber operations. Officials worry that unrestricted public access could give foreign adversaries or even domestic bad actors tools that outpace defenders.
The benchmark context: OpenAI’s existing GPT-5.5-Cyber scored 85.6% on CyberGym — a UC Berkeley benchmark that tests AI agents against 1,507 known software vulnerabilities — outperforming Anthropic’s Mythos 5 at 83.8%. GPT-5.6 is expected to push those scores higher.
- The Anthropic Parallel: Why OpenAI Chose Coordination Over Confrontation
Anthropic’s experience serves as a cautionary tale and OpenAI appears to have studied it carefully. Anthropic released Fable 5 and Mythos 5 publicly on June 9, 2026. Just three days later, the US Commerce Department issued an emergency “Is Informed” letter under the Export Administration Regulations (EAR), requiring an export license before any foreign national could access either model. Anthropic, unable to verify user nationality at scale, disabled both models globally including for its own foreign-national employees within hours.
OpenAI’s strategy with GPT-5.6 is deliberately different. The company was reportedly working with the administration on the upcoming release before the Anthropic incident. By coordinating in advance, OpenAI appears to be pursuing a path toward eventual broad deployment rather than a forced shutdown.
- The Regulatory Context: Trump’s June 2 AI Executive Order
The White House’s request to OpenAI did not happen in a vacuum. Earlier this month, President Trump signed an executive order titled “Promoting Advanced Artificial Intelligence Innovation and Security” — a landmark shift from the administration’s previously hands-off stance toward AI governance.
The order directs federal agencies, including the Defense Department, CISA, and NSA, to develop a voluntary framework for pre-release government testing of frontier AI models. Under the framework, AI companies may submit their most powerful models for government evaluation up to 30 days before releasing them to trusted partners. Critically, the order bars the creation of any mandatory licensing requirement making government review a request, not a rule.
The Fable episode shows the need for clear regulations. Right now, you have an ad hoc, personalized, opaque, possibly lawless approach.”
— Brad Carson, Public First, bipartisan pro-AI safety super PAC, via CNN
- A Chronology of Frontier AI Regulation in 2026
- MAY 28, 2026
Anthropic Announces Claude Mythos Preview via Project Glasswing
Anthropic reveals Mythos Preview, restricted to a small number of trusted organizations through Project Glasswing, citing the model’s extreme cybersecurity capabilities. - JUNE 2, 2026
Trump Signs AI Executive Order
President Trump signs “Promoting Advanced Artificial Intelligence Innovation and Security,” creating a voluntary 30-day government pre-review window for frontier models. An earlier draft proposed 90 days; it was cut over innovation concerns. - JUNE 9, 2026
Anthropic Releases Fable 5 and Mythos 5 Publicly
Both models launch commercially, priced at $10 input and $50 output per million tokens. Fable 5 carries cyber safeguards; Mythos 5 is restricted to Project Glasswing partners. - JUNE 12, 2026
Commerce Department Forces Fable 5 & Mythos 5 Offline
An “Is Informed” letter arrives at Anthropic at 5:21 PM ET, invoking the Export Administration Regulations. Both models go dark globally within hours, affecting users in Africa and worldwide. - JUNE 22, 2026
OpenAI Launches GPT-5.5-Cyber Under Controlled Access
OpenAI rolls out its cybersecurity model which scores higher than Mythos 5 on CyberGym — via its Daybreak program, with prior government coordination and restricted access for verified security professionals only. - JUNE 25, 2026
White House Asks OpenAI to Limit GPT-5.6 Launch
Sam Altman tells staff via memo that the government will approve GPT-5.6 access “customer by customer” during a preview period, with a potential broader release weeks later.
What Comes Next
Altman’s memo suggests that a broader GPT-5.6 release is OpenAI’s goal, the preview period is a means, not an end. If the controlled rollout passes without incident, a wider launch may follow within a few weeks. However, “without incident” is doing a lot of work in that sentence. The government now has an expectation of involvement in the release timeline of any model deemed comparable to Mythos-class capability. That expectation does not easily go away.
Meanwhile, the frameworks themselves are still being built. The executive order gave relevant agencies 60 days to develop a classified benchmarking process for identifying “covered frontier models.” That deadline falls in early August 2026. Until then, decisions about which models require review, and under what standards, remain largely informal guided by conversations between company executives and White House officials, not published rules.
OpenAI said it made clear to the US government that customer-by-customer approval is “not our preferred long-term model.” The industry, and the world, will be watching whether a more sustainable approach actually emerges or whether this controlled rollout becomes the template for every frontier launch that follows.
KEY TAKEAWAYS
* The White House’s Office of the National Cyber Director and Office of Science and Technology Policy asked OpenAI to limit the GPT-5.6 rollout to government-approved partners.
* CEO Sam Altman told staff the government will approve access “customer by customer” during an initial preview window a historically unusual step for a commercial AI release.
* OpenAI and the administration agreed the model’s capabilities are roughly on par with Anthropic’s already-restricted Mythos 5.
* This follows Anthropic’s forced offline of Fable 5 and Mythos 5 on June 12, 2026, after a Commerce Department export control directive.
* A broader public release could follow within a few weeks if the preview period proceeds without incident.
Frequently Asked Questions
1. Why did the White House ask OpenAI to slow-roll GPT-5.6?
The White House’s Office of the National Cyber Director and Office of Science and Technology Policy asked OpenAI to limit GPT-5.6’s release because the model’s advanced cybersecurity capabilities were deemed on par with Anthropic’s restricted Mythos 5. The administration wanted to approve access “customer by customer” during an initial preview period to assess risks before a broad public rollout.
2. What is GPT-5.6 and when will it be publicly available?
GPT-5.6 is OpenAI’s latest frontier model. Because of the White House request, its launch is currently limited to a select group of government-approved partners. CEO Sam Altman indicated that if the preview period goes well, a broader public release could follow within a couple of weeks though no official public date has been confirmed.
3. How does this compare to what happened to Anthropic’s Fable 5 and Mythos 5?
Anthropic’s Fable 5 and Mythos 5 were forcibly taken offline by a Commerce Department export control directive on June 12, 2026, affecting all foreign nationals globally. OpenAI’s situation is different the White House made a voluntary request rather than issuing a legal order, and OpenAI agreed while coordinating with the government in advance. That proactive coordination appears to be the key difference in outcome.
4. Does Trump’s AI executive order make the GPT-5.6 delay mandatory?
No. The June 2, 2026 executive order is explicitly voluntary. It asks AI companies to voluntarily submit their most powerful models for up to 30 days of government review before public release, but it explicitly bars mandatory licensing or preclearance requirements. OpenAI’s compliance is a strategic choice not a legal obligation likely shaped by the lessons of Anthropic’s forced shutdown.
5. Is OpenAI’s GPT-5.5-Cyber more dangerous than Anthropic’s Mythos 5?
On the CyberGym benchmark, GPT-5.5-Cyber scored 85.6% outperforming Anthropic’s Mythos 5 at 83.8%. Yet Anthropic’s models remain offline under a legal export ban, while OpenAI’s model is operating under coordinated, restricted access. Experts have flagged this asymmetry as evidence that the current framework lacks transparent, consistent standards for deciding which models trigger regulation.