The best AI now needs government approval: what GPT-5.6's lockdown means for your business

This week, OpenAI shipped its most powerful model yet. Then it locked the door behind it.

GPT-5.6 is real, and by every account it is very good. You just cannot have it. Neither can almost anyone else. At the Trump administration’s request, OpenAI put its new model into a limited preview, available only to a small set of partners whose names are shared with the government. The most capable AI ever built launched, and the public was not invited.

That is new. And it is worth two minutes of your time, even if you have never typed a word into ChatGPT.

What actually happened

OpenAI announced GPT-5.6 as three models: Sol, the flagship and the strongest of the bunch, Terra, a balanced everyday model, and Luna, a faster and cheaper option. Then, instead of the usual “available to everyone today” launch, the company said the government had asked it to hold back.

In OpenAI’s own words from its Friday post: “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.”

Translation: even OpenAI is uncomfortable with this.

This did not come out of nowhere. Earlier this month the same administration forced Anthropic to pull its Fable 5 model offline over national-security concerns, and effectively banned its Mythos 5 model. The stated worry is cybersecurity: these models are getting good enough at hacking that Washington would rather not hand them to the entire internet at once.

In plain English: the gate moved

For years, the catch with frontier AI was price. The best models cost more per use than the cheap ones, so you picked what you could afford. Annoying, but fair. Money has always been the gate.

The gate just moved. It is no longer only “can you afford it.” Now it is also “are you on the list.” That is a different kind of limit, and a more uncomfortable one. You cannot out-hustle a permission list.

Here is the part nobody is saying out loud: when the most powerful tools go to a pre-approved few first, the gap between the people with early access and everyone else stops being about budget and starts being about who you know. That is not a market. That is a velvet rope.

What this means for you

Let’s be honest, because most coverage will not be: your business does not need Sol. The frontier and the useful are two different things. The models you can already buy today, the ones in Microsoft 365, ChatGPT Team, Claude, and Gemini, are more than enough to draft your emails, clean up your spreadsheets, and answer your customers. Losing access to one experimental model changes nothing about your Monday.

So this is not a reason to panic. It is a reason to pay attention, for three practical takeaways:

• Don’t build your business on a single model you might lose. If one frontier model can be pulled offline by a phone call from Washington, do not architect your operations so that losing it breaks you. Keep your workflows model-agnostic. The tool should be swappable.
• The open and self-hostable options just got more interesting. When access to the top closed models gets political, the models you can actually run yourself become a hedge, not a hobby. We track exactly how good those have gotten in our open-source LLM rankings, and how to run a capable one on your own hardware.
• Capability is outrunning the rules. The reason for the lockdown is that these models are getting genuinely powerful, fast. That is the real headline. The same capability that worries Washington is the capability that, used safely, can take real work off your plate. See what an AI agent can actually do when you point it at the boring stuff.

The verdict

Not a crisis. A warning shot. Gating the best intelligence to a chosen few is a real cost, and you are right to dislike it. But for a normal business, the move that matters is not chasing the model you cannot have. It is getting fluent with the very good ones you already can, and keeping your setup flexible enough that no single company, or government, holds your off switch.

The era of every frontier model landing in everyone’s lap on day one may be ending. The era of AI being genuinely useful to your business is not. Don’t confuse the two.

When you are ready to put this stuff to work without betting the company on any one tool, that is what BlueFort IT does for a living. And if you want the firehose decoded like this twice a week, the verdict included, the newsletter signup is right below.