OpenAI on Friday released its most powerful AI models to date — and then, at the request of the US government, declined to let most of the world use them.
The company unveiled three new systems: GPT-5.6 Sol, its flagship; Terra, a balanced mid-tier model; and Luna, a faster, lower-cost option. Sol is the strongest of the trio, with what OpenAI describes as substantial gains in coding, biology and cybersecurity. But rather than ship them broadly, OpenAI said it would comply with a Trump administration request to limit the initial rollout to a "small group of trusted partners" whose participation, the company noted, "has been shared with the government."
The move is unprecedented. According to reporting from CNBC, TechCrunch and Fortune, it is the first time the US has preemptively asked an American AI company to restrict a launch before it happens. The request was routed through the Office of Science and Technology Policy (OSTP) and the Office of the National Cyber Director, and according to multiple accounts the government will be "approving access customer by customer during this preview period" — a list said to number roughly 20 organizations.
What the government asked for
The administration's intervention did not come out of nowhere. It follows President Trump's June 2 executive order, which asks certain AI developers to voluntarily submit their most advanced models for government review up to 30 days before release. Because the formal vetting framework that order envisions does not yet exist, OpenAI characterized the staggered rollout as a stopgap. The company called the preview a "short-term step" that would put GPT-5.6 on a path to broader availability "in the coming weeks," as it works with the administration on a new cybersecurity framework and a "repeatable process for future model releases."
The timing is also entangled with OpenAI's chief rival. Earlier this month, after Anthropic released its powerful Fable 5 model, the administration ordered the company to cut off access for any foreign national, prompting Anthropic to pull Fable 5 and its Mythos system entirely. OpenAI and the administration reportedly view GPT-5.6 as roughly "on par" with Mythos, particularly on cybersecurity — the capability that appears to have triggered Washington's caution. OpenAI says all three GPT-5.6 models crossed its internal "High" cyber threshold on capture-the-flag testing, with Sol scoring 96.7%, Terra 91.84% and Luna 85.19%.
OpenAI pushes back
OpenAI complied, but it did not pretend to be happy about it. "We don't believe this kind of government access process should become the long-term default," the company wrote in a blog post. "It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."
OpenAI stressed that Sol was built to be safe by design — hardened against adversarial attacks, optimized to favor defensive cybersecurity over offensive exploits, with guardrails baked into the model's behavior rather than bolted on as an external filter. That approach appears aimed squarely at avoiding the backlash that hit Anthropic, whose Fable 5 silently rerouted high-risk prompts to an older model, generating false positives and user frustration. On pricing, Sol costs $5 per million input tokens and $30 per million output tokens, Terra half that, and Luna $1 and $6 respectively — pricing that only matters once general availability arrives.
The precedent, weighed both ways
The core question this raises is how much power the government should have over when, and to whom, a frontier model ships — and reasonable people land on opposite sides.
The case for caution is straightforward: frontier systems that materially advance cyberoffense or bioengineering capability are dual-use, and a brief, vetted preview lets defenders and officials assess real-world risk before bad actors can. Many experts agree the government belongs in conversations about AI and national security.
The case against is about process, not principle. Critics warn the current approach is improvised and opaque. "This escalation of government intervention is nothing to celebrate," said Brad Carson of Public First, describing an "ad hoc, personalized, opaque, possibly lawless approach." Neil Chilson, head of AI policy at the Abundance Institute and a former FTC chief technologist, warned that "continued arbitrary, unexplained deployment of export control authority will make companies slow-walk new models, depriving the public of powerful new tools." Dean Ball, a former White House AI advisor now joining OpenAI, has argued that the June 2 order has created a "de facto involuntary licensing regime" for frontier AI — and that without clearly defined safety standards, the result could be open-ended delays that hand an advantage to China and jeopardize the billions flowing into US AI infrastructure.
What to watch
The near-term test is whether OpenAI's "coming weeks" timeline to general availability holds, or whether GPT-5.6 lingers in vetted-preview limbo the way Anthropic's models vanished altogether. The larger test is structural: whether OSTP and the National Cyber Director convert this one-off into the "repeatable process" OpenAI says it wants — with published standards, defined timelines and a clear legal basis — or whether pre-release gating of frontier models settles in as an informal norm enforced model by model, customer by customer.
"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, Company blog post, 'Previewing GPT-5.6 Sol'