The most important thing about GPT-5.6 is not what the model can do. It is who decided you cannot use it yet. OpenAI's newest and most powerful model launched this week, and for the first time in the company's history, the launch was not OpenAI's call to make. The White House asked OpenAI to limit GPT-5.6 to a small group of partners, and to let the government approve who gets access one customer at a time. OpenAI agreed. So the most capable model OpenAI has ever built is now being released on a list the federal government controls. This is new. Two weeks ago, the government forced Anthropic to pull Mythos and Fable offline after the fact. This time it shaped the launch before it happened. That shift, from reacting to a model after release to controlling who gets it before release, is the real story, and it is bigger than GPT-5.6.

The specifics matter, because this is the first time it has happened this way.

The request came from two White House offices: the Office of the National Cyber Director and the Office of Science and Technology Policy. Commerce Secretary Howard Lutnick also discussed the model directly with Sam Altman before the launch. The ask was straightforward and unprecedented: do not release GPT-5.6 broadly. Instead, limit it to a small group of trusted partners, and let the government approve access customer by customer during the preview period.

OpenAI complied. The GPT-5.6 lineup includes three models, Sol, the most powerful, Terra, a balanced option, and Luna, built for speed and low cost. All three are restricted. Access is limited to partners whose participation has been shared with and cleared by the government. Altman told staff that if the limited release goes smoothly, OpenAI hopes to open it up more broadly within a couple of weeks. For now, the most advanced model OpenAI has ever shipped is available only to organizations Washington has personally signed off on.

The stated reason is cybersecurity. The government considers GPT-5.6 to be on par with Anthropic's Mythos, the frontier model whose vulnerability-finding power triggered the export controls earlier this month. A source familiar with the decision said the government intervened specifically because GPT-5.6 showed Mythos-level capability, and that this scrutiny is simply what now happens with models of that caliber.

Here is the part worth sitting with. There is no law behind any of this.

In early June, President Trump signed an executive order asking AI companies to voluntarily submit their most advanced models for government review thirty days before release. The key word was voluntary. It was framed as a request, not a licensing requirement, and the order itself said so. But in practice, what just happened with GPT-5.6 does not look voluntary. The government asked, and a launch decision moved out of OpenAI's hands. The framework that order promised, the actual testing protocol, the benchmarks, the rules for who decides and on what basis, still does not exist. The review process is being run before the rulebook for it has been written.

That leaves the entire AI industry in a strange place. The most powerful models in the world are now being gated by the government, customer by customer, with no statute defining when an agency can do this, what standard applies, or how a company appeals. Even OpenAI, while complying, pushed back publicly. The company said plainly that it does not believe this kind of government access process should become the long-term default, because it keeps the best tools from the users, developers, enterprises, cyber defenders, and global partners who need them. When the company benefiting from being one of the chosen few is the one warning against the system, that tells you how unusual this moment is.

There are two real sides here, and both deserve to be stated clearly.

The case for the government's approach is genuine. These models are not ordinary software. A model that can find and exploit critical vulnerabilities faster than any human is a dual-use capability, defensive in the right hands, dangerous in the wrong ones. The intelligence community has warned that frontier AI could reshape the cyber threat landscape within months. If you believe that, then making sure a Mythos-level model is not handed to everyone, everywhere, on day one is a reasonable precaution, not an overreach. Better to gate it and get it right than to release it and find out.

The case against is just as real. The government is now shaping which AI products reach the market, with no law, no published standard, and no clear process. Power exercised without a defined framework is power exercised by discretion, and discretion can be applied unevenly. If Washington later clears GPT-5.6 for full release while Anthropic's Mythos and Fable stay restricted, the contrast will look less like consistent enforcement and more like one company getting a break another did not. Anthropic's own objection was never that the government acted, but that no statute defines when an agency can pull or gate a model. The company has called for a clear, transparent legal process grounded in technical facts, rather than decisions made case by case behind closed doors.

Both things are true at once. The security concern is legitimate, and the absence of any rulebook is a real problem. That is the uncomfortable place the industry is in right now.

OpenAI launched GPT-5.6, its most powerful model yet, but the White House asked the company to limit access to a small set of partners and to approve who gets in customer by customer, and OpenAI agreed. The request came from the Office of the National Cyber Director and the Office of Science and Technology Policy, with Commerce Secretary Howard Lutnick weighing in directly. All three GPT-5.6 models, Sol, Terra, and Luna, are restricted, with broader access possibly coming within weeks. The stated reason is that GPT-5.6 has Mythos-level cybersecurity capability. What makes this bigger than one model is the precedent: for the first time, the government shaped a frontier model's launch before release rather than reacting after, and it did so with no law behind it. President Trump's June executive order called such reviews voluntary, but in practice launches are now being directed, and the promised testing framework still does not exist. Even OpenAI, while complying, said this process should not become the default. The security concern is legitimate and the missing rulebook is a real problem, both at the same time. Two weeks ago the question was whether Washington would switch off a model. Now it is who decides who gets one, and the answer, for the moment, is the government.

Stay building. 🤖