Geofenced Out of the Frontier: Why I Run My Own Models
In one week, two labs shipped their best models and drew a border around them. OpenAI's GPT-5.6 preview went to a small set of US partners at the government's request. Anthropic's general-purpose model was switched off and is still offline; the one Anthropic model now cleared to run is a cybersecurity model limited to US organizations that defend critical infrastructure. I build from outside the US, which means the most capable tools on the planet are now ones I cannot legally run. I am not shocked, and I am not going to be bitter about it. I have been building for this case for months.
In a single week, two of the strongest AI labs shipped their best models and drew a border around them. The border is not technical. It is legal, and it runs right through the middle of who gets to use the frontier.
I build from outside the US. That used to be a footnote. This week it became the whole story.
The week the frontier got a border
On June 26, OpenAI introduced GPT-5.6 as a limited preview: Sol, the new flagship, plus Terra and Luna for cheaper, higher-volume work. The models did not ship to everyone. They went to a small group of trusted partners inside Codex and the API, “at the request of the U.S. government,” with general availability promised “in the coming weeks.” Today, June 27, Anthropic said the US government had cleared Mythos 5, which it calls its strongest cybersecurity model, for deployment to a set of US organizations that operate and defend critical infrastructure. Mythos 5 was never broadly available: only a small set of companies, under a program called Project Glasswing, ran it at all. Fable 5, the guardrailed version of that model, is the one the public actually had, and it is the one that got shut down, early this month, with Anthropic negotiating since June 12 to bring it back. It is still dark.
Read those two announcements next to each other and the shape is clear. The constraint on the most capable models is no longer how good they are or what they cost. It is who you are and where you sit. OpenAI will sell Sol at five dollars per million input tokens and thirty per million out, and I would pay that today. The block is not the invoice. It is the passport.
Geofenced AI is here. Mythos 5 is back, but only for select US orgs.
The argument for the fence, and where it breaks
That tweet was the reflex. The considered version starts by giving the other side its best case, so here it is. A model that a lab calls its strongest cybersecurity model is, by the same measure, its strongest cyber-offense model: the capability that finds and patches a flaw is the capability that finds and weaponizes it. Frontier models are dual-use in the most literal sense the term has carried, so a government taking an interest in who runs the most capable one is not, on its face, irrational. That is the strongest version of the argument, and I wanted it on the page in full before I say why I do not buy the result.
Because look at what actually happened. Mythos 5, the cybersecurity model, was never broadly available: only a small set of vetted companies, under Project Glasswing, ever ran it. Fable 5, the guardrailed version the public had, is the one that went dark early this month. That is the mechanic worth staring at. They kept the raw model running for a short list of approved companies and switched off the safety-wrapped version that ordinary people depended on.
I understand the fear. I do not accept that it justifies pulling a public model out of everyone’s hands, or letting the capable ones back only through a US allowlist, and I am not going to call the precedent fine because the press release stayed polite. Even OpenAI, in the same post announcing its limited preview, says it does not believe this kind of government access process should become the long-term default. The companies living under this do not want it either. A frontier lab just lost control of its own launch to a government order, and from outside the US, the way back in is somebody else’s to grant.
What it looks like from outside the border
From outside the US, “available in the coming weeks” is a sentence with no date in it. When the OpenAI preview landed, I quote-tweeted it with the same frustration:
The next leap models are here, but they are also restricted by US export regulations.
Most developers outside the US will not even have the opportunity to test them. The most advanced AI tools are now being gated by compliance, not by their actual capabilities.
Welcome to the era of geo-fenced AI models.
That was the hot take, and half the timeline was posting some version of it that day.
The fuller story complicates that tweet, and in my favor. OpenAI’s own write-up says Sol does not cross the Cyber Critical threshold in its Preparedness Framework; it found the building blocks of an exploit but not a working one, and it is staging the release out of caution about a capability jump it cannot fully bound. So the lockout is precaution turned into process, and the process has no entrance for someone outside the US.
A tweet is where the complaint stops. The work starts somewhere else. So the more useful question, the one I have actually been answering with code, is this: if the frontier can be switched off for me by a regulation I have no vote in, what do I build on instead?
Why I already went local
I run my models on a Mac Studio on my own desk. A local request router I wrote, Jano, sits in front of the model server and folds a burst of mixed requests into a single model swap. The model it serves is a quantized Qwen Mixture-of-Experts that does real work at around 70 tokens a second; I measured the tradeoff against the dense version and made the MoE my daily driver. Last week I gave the router its own telemetry so I can see what the box is doing without anyone’s dashboard. None of that was built as a protest. It was built because owning the inference path pays for itself in cost, latency, and privacy, and the geofence just promoted one more of those dividends to the top of the list.
The model on my desk cannot be revoked. There is no partner list it can fall off, no export tier it can drop out of, no “coming weeks.” It runs at 3 a.m. on a holiday with my internet down. For a solo builder, that property has quietly gone from nice to load-bearing.
It is also why I have been building a mission-control app for self-hosted model servers that I can run from my phone. If the box on my desk is going to be my frontier, I want to operate it from my pocket: watch the queue, swap the model, restart a stuck backend, all from outside the house. The geofence is what turned that from a weekend project into something I want to depend on.
The honest gap
I am not going to tell you a quantized Qwen on a Mac Studio is GPT-5.6 Sol. It is not. The frontier is the frontier for a reason, and on the hardest reasoning and the longest agentic chains the distance is real and visible. The honest framing is a quieter one: for a large and growing share of actual work, the open models I can run are already good enough, and “behind but mine” beats “frontier but forbidden” on every task where the deciding factor is whether the model answers at all. The gap is also closing from the open side faster than the frontier is pulling away. Every month, the model I can run does more of what last month still needed the model I cannot.
What the geofence actually changed for me
The plan did not change; the case I make for it did. Local-first used to be a preference I argued on cost and latency, the kind of thing a reasonable person could talk me out of. Now it is supply-chain resilience: the part of the stack nobody else can switch off is the part worth building on. Last month that was taste. This week it is just true.
The announcements did me a favor. They drew a hard line between the tools you rent and the tools you own, loud enough that nobody can squint past it. When the border went up, I checked my own infrastructure and found most of it already on my side of the line, mostly because the box was the better deal on speed and cost long before it was the safer one.
What I am watching
A few things will tell me which way this settles. Whether “the coming weeks” turns into a real date for developers outside the US or stays indefinite. Whether the cyber Executive Order framework both labs referenced hardens into permanent geography-based tiers or stays a one-time staging step. Whether the open-weight models keep closing the gap fast enough that the frontier’s border stops mattering for everyday work. And whether other labs follow OpenAI and Anthropic here or treat global availability as an advantage worth keeping. If the border hardens, the case for local stops being mine alone. It becomes the default for everyone outside the fence.