Frontier AI Is No Longer a Product, It’s a Permission
Government AI restrictions are policies that allow state agencies to decide which organizations may access the most advanced AI models, turning frontier capabilities like GPT-5.6 from ordinary cloud products into controlled technologies that require case-by-case approval and can be withheld, delayed, or revoked based on security and export-control concerns rather than market demand alone. The key takeaway is blunt: model access control has moved from vendor terms of service into the hands of policymakers, and enterprises can no longer plan AI roadmaps assuming that the latest systems will be broadly available once announced. Instead, they must treat GPT-5.6 availability and similar launches as regulated infrastructure, subject to opaque screening processes that prioritize cybersecurity and biosecurity risk ratings over innovation speed.
On June 26, OpenAI launched GPT-5.6, a three-tier model family it describes as its strongest system to date. Sol targets complex scientific and security tasks, Terra is a mid-tier everyday model, and Luna is a faster, lower-cost option positioned for high-volume work. Yet none of these are generally available: access is initially limited to a select group of US-based partners at the government’s request. For everyone else, the question is not when the model ships, but whether they will be allowed into the gatekeeping process that now sits between AI vendors and their customers.
How Washington Took Control of GPT-5.6 Access
The most striking change is that the government, not the vendor, now defines who is trusted enough to use frontier AI. Twenty companies can currently use GPT-5.6 Sol; every other organization must route an access request through OpenAI into Washington and wait while two White House offices decide whether to say yes. There is no public waitlist and no self-service enrollment, turning model access control into a discretionary approval regime rather than a commercial decision. OpenAI previewed Sol, Terra, and Luna to authorities ahead of launch and agreed to begin with a limited preview for a small group of trusted partners whose participation has been shared with the government.
This arrangement is grounded in both technical risk and policy leverage. Under OpenAI’s Preparedness Framework, Sol earned “High” capability ratings in cybersecurity, biological risk, and chemical risk, and scored 96.7% on internal offensive security benchmarks, crossing the company’s threshold for needing “appropriate safeguards” before deployment. US cybersecurity agencies used those scores to justify gating access, and OpenAI is complying under an earlier Defense Department agreement that lets the Pentagon use its models and gives it influence over distribution. Quotable: “Washington would approve access customer by customer during this preview period,” Sam Altman told staff. That opacity is the mechanism: a gate exists, but the rulebook is not public.
A New Precedent: Restricted AI Models as Security Tools
GPT-5.6 is not an isolated case; it is the second clear data point in a pattern. Earlier this year, Anthropic launched its Mythos-class systems through an invitation-only program limited to a small set of cybersecurity firms and operators of critical infrastructure, after concluding that Mythos exceeded prior models in autonomous cybersecurity tasks. When Anthropic later released Mythos 5 and Fable 5, the US government directed the company to block foreign nationals from accessing the models, citing national security concerns. Shortly after, export controls forced Anthropic to revoke access to Mythos and Fable entirely, and the models have not yet returned. Tech outlets have noted that GPT-5.6 is receiving identical treatment, making government-gated frontier AI launches two for two.
This matters because it sets a working precedent: restricted AI models are now treated like dual-use security tools whose offensive capabilities trigger direct government intervention. The US government’s current approach combines limiting who can access the most powerful systems with punishing foreign actors who try to replicate them through other means, including potential sanctions against overseas AI firms accused of extracting model capabilities via fraudulent API queries. In practice, that means frontier AI is sliding into the same category as sensitive cyber weapons and advanced biotech, even though most enterprises still think of it as SaaS. The shift is quiet, but it is already reshaping how AI supply chains are governed.
Enterprise Reality: AI Roadmaps Now Depend on Policy, Not Product
For ordinary users and most enterprises, the impact is immediate: despite the launch, all three GPT-5.6 models are in limited preview and none are available to the general public. Employees of approved organizations outside the US can use the systems, but everyone else is locked out. OpenAI says the models will become available to the wider public in the coming weeks, and Altman has told staff he hopes for a broader release “a couple of weeks later,” which would mirror the Mythos timeline and point toward late August. Yet there is no hard date, and the government has not stated what criteria would lift the restriction.
That uncertainty forces enterprises to rethink procurement. Terraforming AI strategies around “we’ll adopt the newest model when it launches” is now naive. For every company outside the approved list, the practical question is not when they can access Sol—it is who in Washington controls when they are allowed to ask. Pricing will matter if access is granted: Sol is listed at USD 5 (approx. RM23) per million input tokens and USD 30 (approx. RM138) per million output tokens, Terra at USD 2.50 (approx. RM11.50) input and USD 15 (approx. RM69) output, and Luna at USD 1 (approx. RM4.60) input and USD 6 (approx. RM27.50) output. But the more strategic question is whether AI spend can be planned at all when the most capable models are subject to real-time policy decisions.
The New AI Supply Chain: Policy-Aware, Multi-Model, Less Dependent
These restrictions are already distorting the AI supply chain. Vendors are racing to reduce their dependence on general-purpose chips and to scale custom hardware that can run frontier models in giant data centers, illustrated by OpenAI’s Jalapeño inference chip announcement with Broadcom, aimed at lowering cost per token for workloads like Sol. Yet hardware progress does not help an enterprise that cannot clear the access gate. Instead, smart buyers will start diversifying their model stack, mixing approved frontier systems, less capable but unrestricted models, and open-source alternatives to avoid single-vendor or single-government choke points.
Strategically, enterprises should assume that government AI restrictions and model access control regimes will tighten, not relax, as capability ratings climb in cybersecurity and biosecurity. They will need internal playbooks for what happens when a restricted AI model is suddenly pulled, as Anthropic’s Mythos and Fable were after export controls. The smart move now is to treat AI like regulated critical infrastructure: build policy awareness into procurement, negotiate contingency rights, and design architectures where no single restricted AI model can halt operations if Washington—or any other capital—decides to turn off the tap.
