THE PRIVATE FRONTIER

When the government ordered Fable 5 switched off on June 12, it went dark for me, for my clients, for every developer with an API key. It did not go dark inside Anthropic.

The teams building the new safety classifier spent those nineteen days working with the model. The government researchers who tested the result had access. The Commerce Secretary described two weeks of analysis and approval. All of that work required the thing the rest of us had lost.

I wrote two articles about the ban while it was on. The more I sit with the episode, the more I think I was looking at the wrong part of it.

The interesting fact is not that the model was off. It is that it was never off for everyone. The ban did not switch off intelligence. It narrowed who could stand near it, and the people who kept standing near it were exactly the people already furthest ahead.

This was true before the ban and it is true now. Every model we use is a trailing snapshot of the frontier. By the time a model is released, the lab has been living with it for months: internal checkpoints, versions without production safeguards, successors already in training. The released product is the frontier as it stood some time ago, packaged for the rest of us.

Inside the labs the edge is simply present, all day, in every terminal.

And the edge is not idle. Labs use their frontier models to build the next ones. Anthropic has said that the large majority of its code is now written by Claude. Model-assisted research, model-written evaluations, model-generated training data. The tool works on its own successor, which means whoever has the best tool earliest compounds fastest.

The nineteen days made the resulting ladder visible. The lab kept access the whole time. The government got test access, plus a promise of early access to whatever comes after Fable, as part of the arrangement that brought it back. Mythos 5, the same weights with fewer safeguards, returned on June 26 for a small set of approved American organizations. The general public came back last, on July 1, behind a new classifier.

Four rungs, sorted by proximity to the frontier. In normal times the ladder is hard to see because the bottom rung eventually gets served, and we experience the trailing snapshot as the state of the art. The ban froze the picture long enough to look at it.

Here is what I keep turning over. Export controls, as written, treat access as a switch: this model may or may not be used by these people. But the actual distribution of intelligence is a gradient, and it is steepest at the labs. If capability compounds, the important variable is not only who is allowed to use a released model. It is how large the gap between inside and outside is allowed to grow, and who is watching it.

For my own work the asymmetry is concrete. I consult, write, and build with these models every day, and my edge is a released snapshot plus whatever skill I bring. Someone doing the same work inside a lab has the same skill plus the unreleased edge. In most industries the difference between professional and consumer tooling is convenience. Here the tool is cognition, and it is aimed at producing more of itself.

The labs have a fair answer to all this. Internal access is what safety work requires. You cannot align, red-team, or fix a model you are not allowed to run, and the classifier that brought Fable back was built by people using Fable. The asymmetry is not a conspiracy. It is structural. But structural asymmetries are the durable kind, and this one is attached to the steepest capability curve we have ever had.

I notice that when people argue about when powerful AI arrives, they almost always mean when it is released. Inside the labs it arrives earlier, and lately the interval seems to be growing, with safety testing, staged rollouts, and now government approval added to the pipeline. Each of those steps is defensible on its own. Together they stretch the distance between the frontier and the public.

So the question I am left with is not whether the models will keep getting better. They will. It is whether the rest of us will keep getting them at a delay, and what that delay does when the thing being delayed is intelligence itself.

The frontier is real, and it is currently private. Whatever crosses it next will be used inside first, for months, by people whose access never lapses. The rest of us will read the announcement.