The Strongest AI Models Are Locked Behind a Government Gate. Build Like It Doesn’t Matter.

Early access is a real head start. You can build so it doesn’t decide whether you win.

This month the two strongest models in the world shipped. Anthropic’s Fable 5 went live on June 9 and was gone by June 12, pulled by a government order. OpenAI’s GPT-5.6 never opened to the public at all. It went straight to a list of partners the government helped approve.

The frontier now moves at two speeds. A vetted few get the strongest model first and the time to build on it, while everyone else waits for a later, safer version, by which point the people with contracts and clearances have already turned it into working systems.

You can be angry about that. Plenty of people will be, but their anger won’t buy them a single day of access. The builders who come out ahead quit arguing about whether the gate is fair. They watched it go up and built so it can’t touch them.

The gate’s not going anywhere. So the real work is getting yourself ready for it.

What actually happened

Anthropic released Fable 5 to the public on June 9, its most capable model yet. Alongside it came Mythos 5, the same underlying model with the safety limits stripped off, handed only to a small set of vetted cyberdefense and infrastructure partners. Three days later the Commerce Department sent an export-control order. Foreign nationals couldn’t access either model, including Anthropic’s own foreign-national employees. Anthropic had no way to do that selectively, so it shut both models off for everyone, everywhere.

Then it negotiated. On June 26 the government cleared Mythos 5 to return for more than 100 vetted US organizations that defend critical infrastructure, and Anthropic confirmed it the next day. Fable 5, the version regular people were actually using, stayed dark. As I write this it’s still down, with people close to the talks saying a broader return is coming and no firm date attached.

Same week, OpenAI shipped GPT-5.6, a family it calls Sol, Terra, and Luna. It didn’t reach the public either. OpenAI previewed it to the government first, then opened a narrow preview to partners the government helped vet. Sam Altman told his own staff the government would be approving access customer by customer.

That’s two of the biggest AI companies, same month, both routing their strongest models through the government before the public gets near them. It’s the pattern now, not a one-off.

The reason is cyber, and it’s real

I’m not going to wave this off as a conspiracy, because it isn’t one. The reason these models get gated is specific, and pretending otherwise makes you dumber, not sharper.

Back in April, Anthropic put an earlier version, Mythos Preview, in front of the UK’s AI Security Institute. It solved 73 percent of expert-level hacking challenges that no model could touch a year earlier, and it became the first AI to run a 32-step corporate network attack from start to finish. The same skill that lets a model find and patch a hole in a hospital’s software lets it find and exploit that hole for someone working from a throwaway account. That’s a real capability, and it’s exactly why the government wants a say in who gets through the gate.

So I’ll give the gate its due. Handing the strongest cyber-capable model to anyone with a credit card would be reckless. Somebody had to draw a line. The real question is who gets to stand on which side of it, and whether you can see how the line got drawn.

The head start compounds

A few weeks of early access sounds small when you think of AI as a chatbot that writes better emails. It stops sounding small the second you treat the model as something you build a business around. The first people inside a frontier model get the first learning cycle. They find out which tasks collapse, which workflows break, which jobs the model can quietly absorb, and they wire all of it into systems before anyone else logs in.

Three weeks of that head start counts. Three months counts more. A year of it becomes muscle memory inside an organization, a way of working that’s already humming by the time the same model finally shows up in your dropdown. Everyone eventually gets the same model. The months of head start that came with early access are the real asset.

What the gate can’t reach

The gate has a hard limit, and that’s where the resolve comes from. It controls which model you can open. How good you are with the ones you’ve already got is entirely yours.

Most people are held back by how shallowly they use the tools already in front of them, not by a locked door at the frontier. (I’m in this too. There are about 299 new AI things to learn every day, and life is always life-ing. Kids, business, health, all of it.) That’s a practice problem, and you can start closing the gap it creates today.

The builders who come out ahead make their work portable. They own their data and keep it clean. They write their workflows down instead of leaving them in their heads, and they keep their prompts somewhere other than one company’s chat window. Anything that carries real weight gets human review before it ships. And nothing they’ve built is locked to a single provider, so when one changes the rules overnight (exactly what just happened to Fable 5), they switch to another model and keep moving.

That’s the whole point. Own your systems, stay portable, and losing access to one model becomes a quick switch instead of a crisis.

You don’t need to be first in line

This isn’t a one-time event. The strongest models will keep going to vetted institutions first, and regular users will see them later. You don’t have to like that to be ready for it.

The people who do well from here are the ones who built so they never needed early access. They own their data, keep a clear record of how they actually work, and can move to whatever model is best that week without starting over.

None of that requires permission. You can start today with the tools already on your screen.

I’d rather name this now, while it’s still forming, than explain it to you a year from now when the gap is a lot harder to close.

If you want the next piece, I’m putting together the actual setup, the workflow I’d build right now so it doesn’t lean on one company, one model, or one permission layer. It goes out to subscribers when it’s ready. Subscribe and you’ll get it.