For most of the modern AI era, getting the best model meant one thing: open your wallet. You signed up, accepted the terms, and paid per token. That quiet assumption — that capability flows to whoever is willing to pay for it — is the thing now breaking. The U.S. government will vet who is allowed to use OpenAI's newest model, GPT-5.6, and the same week brought reports that a new Anthropic model was cleared for release only to "trusted partners." Access to the frontier is turning from a purchase into a permission.
What actually changed
Until now, the gate in front of a new model was commercial. Anyone with a credit card and an accepted usage policy could call the most capable system a lab offered. Layering a government review on top of that changes the nature of the gate entirely. It is no longer "can you pay and will you follow the rules?" but "have you been approved?" That is a different question, with a different answer for different people, and it moves the decision about who gets to build with frontier AI partly out of the market and into the state.
Why "vetting" is a bigger deal than it sounds
Vetting implies a list — and lists have insiders and outsiders. The moment access depends on approval rather than payment, the most powerful tools split into two tiers: a public tier anyone can use, and a restricted tier reserved for vetted institutions. For years the defining story of AI was relentless democratization, where each capability jump shipped to nearly everyone at once. A vetting layer quietly ends that. The newest frontier may now arrive late, or not at all, for anyone outside an approved circle.
The export-control playbook, applied to software
There is a clear precedent here, and it is worth naming: export controls. The advanced chips that train and run these models are already restricted, with the government deciding which buyers in which countries may purchase them. Treating the model itself the same way is the logical next step — the model is, after all, the thing those chips were built to produce. The difference is that software is not a crate of hardware you can stop at a port. It is weights, and weights are copyable. Applying an export-control mindset to something that can be duplicated infinitely is going to be far messier than controlling physical goods, and that tension is where most of the coming fights will happen.
Who this squeezes
The people who feel this first are not the giant incumbents — they will be on the approved list. It is the smaller builders: a two-person startup, an independent researcher, a developer in a country that is not a close ally. These are the users for whom "vetted access" most easily becomes "no access." The risk is a frontier that is technically available but practically gated, where the gap between the best model the public can touch and the best model that exists keeps widening. That gap is exactly what an open ecosystem spent years trying to close.
The open-weights counter-pressure
This is also the strongest argument the open-weights camp has had in a while. You cannot vet access to a model you can download. Once weights are public, there is no approval step to clear — the model is simply out there, runnable on your own hardware. Expect the labs and communities betting on open weights to seize this moment, framing themselves as the part of the field that no one can gate. If the closed frontier becomes permissioned, more builders will route around it, and the center of gravity for everyday development could shift toward whatever capable model is freely available, even if it trails the absolute state of the art.
What to watch next
The important questions are all about process, and right now they are unanswered. Who decides who is "trusted"? On what criteria? Is there an appeal if you are turned down, or any way to know why? Vague "trusted user" standards are easy to apply unevenly and almost impossible to contest. The real test is scope. If vetting stays narrow — genuine misuse, clear national-security edge cases — it may be a manageable friction. If it quietly expands into a general permission layer over who gets to build with the best AI, that is a structural change to the industry, and one that would be very hard to reverse once the machinery exists.
Our take
Gating the frontier is being framed as a safety measure, and some narrow version of it may genuinely be one. But access controls have a way of outgrowing their original justification, and a permission layer over capability is the kind of infrastructure that is easy to build and hard to dismantle. The most consequential line in AI may no longer be the one between models that work and models that don't — it may be the one between the people allowed to use the best of them and the people who aren't. Watch where that line gets drawn, and who gets to draw it.
Trending on The Washington Post — analysis by GENZTech.