The US Government Is Choosing Who Gets AI Now

GPT 5.6 shipped this week, and it went out about as quietly as a flagship launch can go. That is not an accident. The US government has restricted access to a very small number of companies, the same move it pulled on Anthropic's Fable and Mythos a few weeks back. So the biggest model OpenAI has ever released is out, and almost none of us can touch it.

There is a lot packed into that. What is actually going on with GPT 5.6 and these restrictions, how China is treating this whole mess as the opening it has been waiting for, why the US is scrambling to stop China distilling its models, and why Coinbase just walked away from frontier models toward open weights out of China to save money. I think we are hitting a genuinely dangerous point in AI, and not only for the reason you would guess. The US lead is suddenly in question, sure. But the way we are choosing to hand out new models inside the US might be the more dangerous half. Let me walk through it.

GPT 5.6, before we get to the scary part

I will keep this short, because the model is not in my hands yet and I do not hand out opinions on things I have not tested. Here is what OpenAI actually showed.

There are three tiers now. Soul, the flagship. Terra, the balanced everyday model. Luna, the cheap fast one. OpenAI says Soul is their strongest model yet, with real jumps in coding, biology, and cybersecurity. Terra is pitched as competitive with GPT 5.5 while being 2x cheaper. A standard tier ladder.

Then comes the line that gave me flashbacks. We believe in broad access, and we plan to make these generally available in the coming weeks. Remember when Anthropic said they would have news in 24 hours? That was June 12. Coming weeks, sure.

The reason for the delay is spelled out right there in the announcement. At the government's request, OpenAI is starting with a limited preview for a small group of trusted partners, whose participation has been shared with the government, before releasing more broadly. OpenAI is careful to add that they do not think this kind of government access process should become the long-term default, that it keeps the best tools away from developers, enterprises, and cyber defenders who need them. Read that again. The company that built the model is publicly saying it does not want to release it this way. It is being made to.

I want to flag one thing on the benchmarks, because it tripped my usual alarm. They are showing GPT 5.6 Terra, the tier they themselves say is on par with GPT 5.5, landing at the same terminal bench score as Fable 5. I tested Fable 5. I made a whole video about it. It did not last long before the rug pull, but I used it, and it was clearly a step above both 5.5 and Claude Opus 4.8, a model in a class of its own. So no, I do not buy that a 5.5-equivalent model matches it on terminal bench. Every lab does this on launch day, Anthropic included. There is always one number in the deck that does not smell right, and this is that number. I will hold the real judgment until I can run my own benchmarks.

The part that is genuinely interesting is the pricing. If the top of this lineup really is Mythos class, and they are pricing it around 5 dollars in and 30 out per million tokens, that is roughly half of what Fable ran, which was closer to 10 and 50. A Mythos-class model at half the price, and apparently token efficient on top of it. If that holds up when I can test it, it is a very big deal for OpenAI. But that is a conversation for launch day, whenever the coming weeks arrive.

The part that actually worries me

Set the model aside. This is the thing.

The US government released Anthropic's Mythos to a short list of US companies. That happened, we all lost access to Fable, Anthropic said it would share details in 24 hours, and then nothing. That was June 12. As I record this it is over two weeks later, and the only thing that has moved is that GPT 5.6 got the exact same treatment. A powerful model, handed to a handful of companies, no sign of a broader release.

Update, July 1: hours after I recorded this, the government lifted the export controls and Fable and Mythos came back, nineteen days after they vanished. Good news, and it changes nothing about the argument. The gate is the release pipeline now, and GPT 5.6 is still sitting in it.

One of the original sticking points was that access was being restricted to US nationals. Think about what that actually requires. You have to verify that every single person you give the model to is a US national. Maybe you can bolt enough KYC onto a subscription to muscle through that. How do you do it on an API? You cannot, not really. I would bet that requirement is a big part of why this has been frozen since June 12.

It got absurd enough that non-US nationals working at Anthropic, some of the leading AI researchers on the planet, could not use models their own company built. That specific piece got relaxed, those employees can use it now, and non-US staff inside the approved companies get access too. But that is a grain of sand in a very wide desert, because the real problem is that this looks like the norm now, not the exception.

Pulling the ladder up

Jeremy Howard, who is as credible as AI researchers get, called the shape of this early, back around Project Glasswing and the Fable release. His line was about silently sabotaging experiments in order to stultify scientific progress and protect a technology lead. He was pointing at Anthropic hand-picking who gets Mythos as a way to protect its own position. Now the government is running the same play at a national level, best models to a select few.

The phrase that keeps coming to mind is pulling up the ladder. If you are not one of the anointed companies, catching up gets a lot harder, because you cannot use the best models for research, and you cannot point them at any kind of recursive self-improvement. You wait until someone decides the public gets a turn. It is brutal for startups and for the whole market sitting underneath the frontier. You get the model when they say so, if they ever say so, and that is before whatever new KYC hoop shows up next to lock out foreign access.

Justin Murphy had the sharpest version of the reaction. If they really start gatekeeping who gets to use the best models, he calls it a declaration of war, and says the prospect fills him with the most sincere cipherpunk will to power he has felt since he was a teenager. His answer, if it goes that way, is to go all in building the most psychotic swarms of open-source models and fine-tunes possible, aimed at jamming the entire institutional public sphere, because otherwise the marketplace of ideas is over before we notice. Strip the theatrics and the point is simple. If the ladder gets pulled up politically, so only select institutional players hold the most capable intelligence, you are being cut out, and you should be as energized about it as the people who get energized about their most protected rights. Translation, prepare. This starts to look like an arms race, and you are not on the invite list.

Meanwhile, China is having the week of its life

Now look outward, because this is where it gets strange. China is very obviously treating this as the opening it wanted.

If you have not seen it, resellers in China are offering Claude tokens at 70 to 90 percent below Anthropic's official API prices. The how, per the breakdown that went around Hacker News, is a mix of things. They pool spare capacity from Claude Max accounts, they lean on payment fraud, and they sell the access anonymously, below cost, at a scale of tens of thousands of bot accounts.

And if that sounds like rumor mill, here is the official version. Anthropic told US senators that actors linked to Alibaba ran roughly 25,000 fraudulent accounts and pulled 28.8 million exchanges out of Claude in about six weeks, aimed at exactly what Claude is best at, agentic reasoning, software engineering, long-horizon tasks. That is a distillation run. You harvest millions of input-output pairs and train your own model on them. It is illegal, it works, and it is how China sits just behind the frontier for a fraction of what the frontier costs to build.

We have watched this movie before. China in the 60s and 70s was badly under-industrialized, then it figured out how to manufacture basically anything the West makes, more efficiently. Look at cars now. The export bans are not going to help as much as people want them to. There are Chinese APIs today selling Claude Opus 4.8 at something like a 90 percent discount, and those are not even the export-restricted models. If they want the restricted ones badly enough, I would not bet on a ban being the thing that stops them.

My honest read is that staying ahead means doing the opposite of what we are doing. Free up the internal market. Let US companies compete hard, let them give the public access, let them stay profitable and pour those profits back into pushing the frontier out further. Right now the export bans mean labs that sank billions into these models cannot sell them to you or me. They can only sell to a few blessed companies. That is not how you fund the next model.

The SK Telecom mess

Here is a concrete example of the bind. The White House decided it could not trust Anthropic to safeguard its latest models, and the trigger was a Korean telecom giant. Anthropic had let SK Telecom, one of the largest companies in Korea, use its Mythos-class models. The White House then told Anthropic to cut them off, because it believed SK Telecom was leaking access to China. Anthropic complied.

The awkward part is that SK Telecom is also an investor in Anthropic. And in its corporate structure it has long-standing, deep ties to China. So I can believe the White House on this one. But it proves the point I keep landing on. This is the age of the internet. Keeping China out with export bans is going to be very hard, and while we try, we are kneecapping our own market and our own labs' ability to stay at the frontier. I am genuinely worried that in trying so hard to protect the technology, we end up doing the damage ourselves.

What the drone industry already taught us

I keep coming back to drones, because we already ran this experiment and we know how it ends.

Consumer drones are completely owned by China, specifically DJI. The US market was big, around 6 billion dollars, and DJI took it by undercutting on price, propped up by subsidies and the Shenzhen supply chain. Same story as EVs, where the government subsidizes the key Chinese players so they win the market outright. The US companies, 3D Robotics and Airware, tried and lost. And notice why they lost. It was not a missing feedback loop, they were not failing to get product in front of customers and iterate. They lost the price and iteration war flat out, because you cannot beat subsidies plus the most efficient supply chain on earth. China ended up with around 70 percent of the global market, Chinese brands became the standard everyone else measures against, and by the time the US started issuing bans, the cost lead was already locked in. The bans went nowhere.

Now map that onto AI. We are over-restricting who is allowed to deploy frontier models, and keeping the cost high, while China ships open models into the world. If the labs cannot recoup their training costs as inference prices collapse underneath them, that is a real problem, and inference prices are collapsing. The timing is not helping either. The whole pitch was that models move quickly from the initial companies to a broad release. It is not moving quickly. It has been stuck for weeks. String enough of that together and China's open models become the global default, at a fraction of the cost, and once that happens it does not get undone. That is the exact lesson the drones left us.

But who would ever switch? Ask Coinbase

The obvious pushback is that US models are just better, so who would actually switch. Coinbase would. Coinbase already did.

Brian Armstrong, the CEO, announced that the company moved to models like GLM 5.2 and Kimi K2.7. The reason was blunt. Token usage at Coinbase had shot up in recent months and the cost had become unsustainable, which is happening at basically every company with heavy AI adoption, whether they are saying it out loud or not. Swapping in the Chinese models, plus some token-maxing on top, cut their AI spend nearly in half. For a company that size, that is a staggering number.

So it is not a someday problem. It is happening now. Companies are waking up to two things at once. Frontier models are expensive, and access to them is unpredictable, because you do not know if or when the government pulls the model out from under you. That second one is a reliability problem, and reliability is exactly what a company like Coinbase cannot compromise on. So they build something cheaper that they actually control. I have no doubt we see a lot more of these announcements over the next few days.

Open weights are not the finish line

Here is where I think this goes, and it is the part I am actually optimistic about.

Open weight is not open source. Better than closed weight, of course. But for the industries where an AI decision carries real consequences, healthcare, autonomous vehicles, financial underwriting, partial transparency is not enough. Full accountability means understanding not just the finished model, but how it was built, what data it was trained on, and the points where it might have drifted from anything resembling ethical best practice. And that is exactly what we are never going to get out of China.

Think about why. China is largely building its models by distilling US models, which is illegal. So they cannot go fully accountable and publish their training data without more or less confessing to it. Add the CCP's own restrictions on top, and full transparency is off the table for them.

Which is the opening. If a company can crack genuinely open-source models, make them profitable, and deploy them to large enterprises at a fraction of frontier cost, that becomes a huge arm of this whole industry, and it is one China structurally cannot fill. Open weights are a step, not the destination.

So here is where I land. The ban strategy needs a rethink. Not a tweak, a real review, one that takes the cost side seriously and takes seriously what China is already doing with US models over the open internet. We are trying to protect a lead by building a wall around it, and the drone industry already showed us how that ends when the other side is willing to ship cheap and iterate. The most valuable thing right now is not the biggest model behind the tightest gate. It is a healthy market that keeps the labs funded and puts the tools in enough hands to matter.

So watch what your favorite companies quietly switch to over the next few weeks. That is the real scoreboard, not the launch-day benchmarks. And right now it is not pointing where the bans want it to.