The US Government Now Decides Who Gets OpenAI's Most Powerful AI First

OpenAI launched its strongest model yet on Friday. Access is controlled not by the company, but by a White House approval list with no published criteria, no formal legal basis, and no deadline.

OpenAI released GPT-5.6, its most capable AI model to date, on Friday. The roughly 20 organizations that can actually use it were not chosen by OpenAI. They were approved, one by one, by the US government.

The launch marks the first time an American AI company has released a frontier model under a government-managed access list. It follows the Trump administration's forced shutdown of Anthropic's Fable 5 and Mythos 5 two weeks ago and arrives on the same day the government partially restored Anthropic's Mythos 5 for a separate list of over 100 approved organizations. The pattern that has emerged over the past fortnight is no longer ambiguous: Washington has inserted itself as a gatekeeper for frontier AI releases, with no formal regulatory framework to govern how that gatekeeping works or when it ends.

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

The company complied anyway.

A Three-Tier Model Family, All Rated High Risk

GPT-5.6 ships as a family of three models. Sol is the flagship, built for demanding tasks including complex coding, long-horizon cybersecurity research, and agentic workflows. Terra steps down on peak capability while delivering performance comparable to GPT-5.5 at roughly half the cost, making it the practical choice for high-volume business use. Luna sits at the bottom of the range on price and is optimized for speed and routine automation.

Sol is priced at $5 per million input tokens and $30 per million output tokens. Terra comes in at $2.50 and $15, and Luna at $1 and $6. On the ExploitBench cybersecurity benchmark, Sol is competitive with Anthropic's Mythos Preview while using approximately one-third of the output tokens, a cost efficiency advantage that matters at the scale at which these models are deployed.

The risk classification across the family is notable. All three models, including the lower-cost Terra and Luna, are rated "High" by OpenAI's internal preparedness framework for both cybersecurity and biological and chemical capabilities. That makes GPT-5.6 the first OpenAI model family where every tier triggered the highest internal risk classification. OpenAI says Sol did not cross its internal "Cyber Critical" threshold and is "better at helping people find and fix vulnerabilities than reliably carrying out end-to-end attacks." The company dedicated roughly 700,000 A100-equivalent GPU hours to automated red-teaming before launch and found what it describes as universal jailbreaks across the model family before any public access was granted.

Sol introduces two new operating modes. A "max" mode gives the model extended time to reason through difficult problems. An "ultra" mode goes further by deploying multiple sub-agents working in parallel, embedding multi-agent coordination directly into the model's behavior rather than requiring external orchestration.

How the Gating Actually Happened

OpenAI had been briefing the White House on GPT-5.6's capabilities for roughly a month before the launch. CEO Sam Altman met with administration officials in early June. The company anticipated that some form of staggered release might be required, but did not expect the government to approve each customer individually and cap the initial cohort at around 20 organizations.

The formal request came from two White House bodies: the Office of the National Cyber Director and the Office of Science and Technology Policy. Their stated goal was to test and evaluate the model's security profile before broader access was granted.

Commerce Secretary Howard Lutnick went further. According to reporting by The Information, Lutnick called Altman directly to warn him against releasing GPT-5.6 to the public without sign-off from additional agencies, even after OpenAI had already briefed senior officials. Lutnick's intervention replays a pattern he established with Anthropic: the same official who issued the June 12 export control directive against Fable 5 and Mythos 5 is now the central figure managing OpenAI's release.

Altman described the situation to employees in an internal memo on Thursday. He said the government would be "approving access customer by customer during this preview period" and that a broader release would likely follow "a couple of weeks later" if the process went smoothly. He also told staff that OpenAI had made it plain to the White House that this model of government oversight is "not our preferred long-term model."

Amazon's Bedrock platform serves as one of the access routes for the approved partners. The current cohort is entirely US-based, though OpenAI said it expects employees of approved companies in "supported countries," including the UK and Australia, to have access to the model.

The Gap Between Voluntary and Mandatory

The administration's legal hook for all of this is a Trump executive order signed on June 2, 2026, which asks AI companies to voluntarily give the government up to 30 days of pre-release access to models with advanced cyber capabilities. The word "voluntary" is doing significant work in that sentence.

OpenAI's situation is technically voluntary. The company chose to cooperate rather than face the coercive alternative that Anthropic experienced. What that means in practice is that the distinction between a voluntary review and a mandatory licensing regime has narrowed to the point where it may no longer be meaningful.

Dean Ball, a former White House AI adviser who is joining OpenAI, put it directly. He argued the executive order has created a "de facto involuntary licensing regime" for frontier AI, made more dangerous by the absence of clearly defined safety standards. Without those standards, he warned, the approval process could produce endless launch delays that hand China an advantage in the AI competition while simultaneously undermining the return on the billions being invested in US AI infrastructure.

The formal framework that is supposed to govern all of this does not yet exist. Federal agencies have an August 2026 target deadline to finalize a benchmarking and assessment process for frontier models. Until that process is established, the ad hoc partner-approval mechanism used for GPT-5.6 is the actual procedure. The executive order's 30-day window is also the compressed version: an earlier proposal called for 90 days and was shelved before the current order was signed.

Brad Carson, head of Public First, described what has emerged as "an ad hoc, personalized, opaque, possibly lawless approach" to AI oversight. He added that while government intervention in dangerous AI deployments is appropriate, it must be "done in a way consistent with transparency and basic fairness."

The Fairness Question OpenAI and Anthropic Cannot Both Answer

The two companies' situations on Friday were sharply different, and the difference is hard to explain on purely technical grounds.

OpenAI launched GPT-5.6 with a clear path to broad availability in "the coming weeks." The government's review is framed as a temporary gate before general release. Anthropic, whose Mythos 5 the government has described as having capabilities comparable to GPT-5.6 Sol, had Fable 5 forcibly shut down two weeks ago and has not received a timeline for its return. Mythos 5 was partially restored Friday for over 100 specific organizations, but Fable 5, the publicly available product built on the same architecture, remains offline for every user on the planet.

A source close to the OpenAI discussions told Axios that the government intervened on GPT-5.6 because it has "Mythos-like" capability, "not because the administration is suddenly taking a heavier hand." The source added: "This is what's happening with models of that caliber."

If that is the standard being applied, the question of why Anthropic's Fable 5 remains dark while OpenAI's equivalent model moves toward broad release has not been publicly answered. Anthropic's public posture has been to challenge the legal basis of the action, calling for "a statutory process that is transparent, fair, clear, and grounded in technical facts." OpenAI's posture has been to comply while signaling disagreement.

The Decrypt and others have raised the concern that if the largest AI companies help design the rules for frontier AI oversight while those same rules are enforced unevenly across the industry, the result could function as regulatory capture: a framework that benefits the labs that helped write it.

A White House That Changed Direction Without Saying So

The current moment is a significant departure from where the Trump administration began.

Last year, Vice President JD Vance said that "excessive regulation of the AI sector could kill a transformative industry." On its first day in office, the administration rescinded a Biden-era requirement for AI companies to submit safety tests to the government before releasing powerful models, calling that requirement overly burdensome.

Nine months later, the same administration is approving commercial AI customers one by one before a model can launch.

The shift happened against a specific backdrop. The UK's AI security body formally assessed Anthropic's Mythos as a "step up" over previous cutting-edge models in terms of cybersecurity capability. The NSA ran an authorized red-team exercise using Mythos against its own classified systems and found vulnerabilities within hours. Those two data points, one from an allied government's official security assessment and the other from inside US intelligence, landed on an administration that had told the industry it would get out of the way.

The industry is now learning what getting back in looks like.

OpenAI said it believes the Trump administration still has US AI competitiveness in mind despite its recent moves. The company also said it plans to work with the White House on a more sustainable approach for future model releases. Neither statement carries a timeline. The formal framework for evaluating frontier models is due in August, but its criteria, its scope, and its enforceability remain undefined. Until those questions are answered, the question of who decides when America's most capable AI reaches the public does not have a legal answer. It has a phone number: Commerce Secretary Howard Lutnick's.