The Frontier That Lasted Three Days

What one model's brief life did to four essays I wrote in winter

I upgraded to the Max plan on a Tuesday, for one reason. Claude Fable 5 had been public for a few days and it was the first model in a while that felt clearly ahead of everything else I could run, not on a chart but in the work, the kind of difference you stop noticing you depend on until it is gone. Three days after launch it was gone. On the afternoon of June 12 the US government sent Anthropic an export-control letter. The order barred any foreign national from using the model, nobody can verify citizenship on a per-request basis, and so Anthropic switched it off for everyone.¹ I live in the Netherlands . For the purposes of the directive, that made me the problem.

I bring this up not because my weekend was disrupted (it was, mildly) but because it folded an argument I’d been having with myself since January into a single week. Between January and April I’d published four pieces that, lined up next to each other, looked like they couldn’t all be right. The Perception Threshold argued that improvement stops mattering once nobody can feel it. In The Token Divide I said frontier access is decisive enough to tier a market by price. Local Parity had open and local models closing the gap until the difference disappears, and Sovereignty as a Structural Force put jurisdiction, not just capability, in charge of who depends on whom. The first two, in particular, seemed to point in opposite directions.

Then June settled it, and not in the soft way essays usually get “confirmed,” where you squint and the trend line agrees with you. Three model releases and one government letter, inside about a week, turned four speculations into a description of something that had already happened.

The divide was real, for three days

In April I argued that as frontier models pull ahead, access to them tiers by price: the strongest model, the highest limits, the longest agentic runs migrate into the plans only some people can afford, and that gap compounds into a difference in how fast you can work. Fable 5 was that argument with a face. It shipped to the public on June 9 at ten dollars per million input tokens and fifty per million output, roughly double the going rate for Opus,² and on the hardest reasoning test in common use it sat about twenty points clear of everything else you could buy.³ Not clear of the weak models: clear of the field, the same low-forties band that Opus 4.8, GPT-5.5, Google’s current public model,⁴ and the best open weights were all crowded into.⁵ I paid for Max because, for once, the divide was something you could feel.

What I got wrong was how it would end. I’d assumed the divide would erode the ordinary way: diffusion, cheaper compute, open weights leaking the capability downhill. Fable 5 never diffused; it was withdrawn.

Sovereignty arrived as a Friday-afternoon letter

In February I’d written that once these models are treated as strategic infrastructure rather than products, sovereignty enters the arithmetic: depending on a single provider under a single jurisdiction becomes a risk that can outweigh a performance edge, and that pressure pushes the industry toward redundancy and diffusion instead of clean consolidation. I framed it as a slow structural force. It arrived late on a Friday afternoon, as a letter.

The instrument was a Bureau of Industry and Security “Is Informed” letter under the Export Control Reform Act, the first time that machinery has been aimed at a commercial AI model.⁶ It required an individually validated license before any foreign national, anywhere, could touch Fable 5 or its ungated sibling Mythos 5, and it swept in Anthropic’s own non-citizen employees. There was no court. The government published nothing and explained nothing, and over the following days the stated reason shifted more than once. What the public knows of the order, it knows because Bloomberg obtained a copy and published the text four days later.⁷ The models are still dark as I write this.

I don’t need to know the real rationale to make the point, since not knowing it is the point. A capability that hundreds of millions of people were using on a Tuesday was unreachable by Friday, by administrative fiat, with effectively no notice. That is what “dependency is a risk variable” means once it stops being a phrase in an essay. And it was not only me. Two days after the shutdown, more than eighty cybersecurity executives and researchers, people at Nvidia, Adobe, Zoom, Sophos, signed an open letter to Lutnick and the National Cyber Director arguing the threat had been overstated and the restrictions should be lifted, and within a week more than a hundred names were on it:⁸ the defenders the order was ostensibly protecting, asking for the capability back. The people who actually ship things read it the same way. Within days the conversation had moved to gateways that can swap one model for another without a rewrite, and to running open weights on hardware you control, not because those models score higher but because a model nobody can cut you off from is worth a discount in raw capability. Single-vendor dependence became, more or less overnight, its own category of operational risk.

Local parity stopped being a forecast

Local Parity, from February, made a narrower claim than the open-source case usually gets credit for. The argument was never that open models would match the frontier outright. It was that they’d close the gap to the point where, for most real work, you couldn’t tell the difference, and that once inferiority becomes imperceptible it stops being an economic fact.

GLM-5.2 landed on June 13, four days after Fable 5 and one day after Fable 5 went dark, which is timing I couldn’t have scripted.⁹ The weights themselves followed on June 16, under a plain MIT license with no regional carve-out. It runs on infrastructure you own, and it costs roughly a sixth of what the closed frontier charges: four dollars forty per million output tokens against Opus 4.8’s twenty-five.¹⁰ On the things people do with these models all day (front-ends, mainstream coding, agentic tool use, competition math) it lands level with the closed field or within a point or two of it, and it sits second on the human-voted coding arena.¹¹ The people who use it daily describe it as the first open model that simply feels right inside a coding harness.¹² It doesn’t win every benchmark, and it doesn’t need to. For the median task and the median user, you can no longer feel which side of the open/closed line you’re standing on. The condition I’d named as a forecast is now a price list.

There’s a tail where the closed field still scores higher: the longest, gnarliest agentic runs, the multi-file refactors that grind for hours. I want to be careful about that lead, because I caught myself overstating it. The closed models post better numbers there.¹³ Whether they are better there is a separate question, and the honest answer right now is that we don’t fully know, because the only evidence is a benchmark layer the field spent this same spring learning to distrust, a stretch in which a three-billion-parameter model running on a laptop posted flagship-level math scores and set off weeks of argument about contamination and gaming.¹⁴ When a lead shows up both on a contested leaderboard and in what experienced people report after weeks of real use, I’ll grant it. Where it shows up only on the leaderboard, I’ll hold it open. Either way the tail is narrow and shrinking, and it is held by closed models living under the same export regime that just made Fable 5 disappear.

Why the two essays were never fighting

Here is where the two pieces that looked like they contradicted each other stop contradicting. The Perception Threshold said improvement quits producing advantage the moment users can’t perceive it, while the claim in The Token Divide was that frontier capability is decisive enough to tier a market around it. Both turned out to be true, because they were describing different people.

For the median user and nearly every ordinary task, the available models have converged into that low-forties cluster I keep returning to. The differences sit below the threshold, you can’t feel them, and the cheaper, more portable, more sovereign option wins. That is local parity behaving exactly as advertised. The token divide lives somewhere else, out at the extreme tail, on the hardest problems, where a Fable 5 could open twenty real points and a heavy user would pay to stay on the right side of the gap. Two regimes in one industry: the threshold governs the middle and the divide governs the edge.

What decides which regime you actually live in is exposure, not capability. The middle is resilient because it is open and diffuse: many providers, swappable, runnable on your own metal. The edge is premium and closed, and now demonstrably revocable. Fable 5 proved the edge can be real, and it proved the edge can be switched off by a government you don’t vote for, on a timescale of hours. Faced with that, even an actor who can afford the edge has a reason to design for the middle. The frontier advantage kept its value and lost its guarantee.

The scorecard

Four months on: the perception threshold held, and held wider than I expected. Local parity arrived on a price list. The token divide was real for as long as Fable 5 existed, and the manner of its disappearance is the cleanest demonstration of the sovereignty argument I could have asked for and would never have wished for. Read together, the four pieces describe one system seen from four sides rather than four predictions, and June ran it as an experiment.

The thing I’ll keep from the week isn’t in any of the four essays, though. It’s a note about evidence. The most reliable data point in all of this was a model going dark, not a benchmark; benchmarks were the part I had to learn to distrust. When I sat down to wire my own setup back together after the weekend, I didn’t reach for the best score. I put a gateway in front of everything so no single model could strand me again, and I pulled GLM-5.2 down onto a machine I own. I built for the middle, which is, I suppose, the most honest validation a person can give his own argument: not citing it, but acting on it.

The floor is on loan

Everything I’ve described, the parity and the choice it gives you to build for the middle, rests on a single fact that has nothing to do with my preferences and everything to do with someone else’s strategy. The reason the token divide is not, right now, a runaway train is that Chinese labs keep flooding the market with frontier-grade open weights under licenses that don’t even bother to carve out their largest users. GLM-5.2 is the example in front of us, but it’s a pattern that runs through DeepSeek, Qwen, Kimi, and now even a nine-person team at a social-media company. That flood is what holds the price floor down. Western open releases help at the margin, and Google’s Gemma line applies real pressure, but the weights that actually sit at the closed frontier’s heels with no strings attached are coming out of China. Remove that supply and the floor goes with it.

And the supply can be removed, by three different routes that arrive at the same place. The first is an outright cartel: illegal, and also a thing that happens. The second needs no conspiracy at all. The ordinary gravity of an oligopoly, what Hotelling described as competitors converging on a single position, would pull the two frontiers toward the same high pricing without anyone making a phone call. The third is the route we just watched work. A government decides a class of models is too strategic to circulate, and the supply stops in an afternoon. The United States did precisely that to Fable 5. Nothing guarantees that Beijing will keep treating its best open weights as an export to encourage rather than an asset to withhold, and the day its labs lead instead of chase, the reason to give the lead away disappears. We have already seen the rehearsal: as Chinese models became good enough to monetize, some of the strongest stopped shipping open.

So the open-weight floor is a byproduct of one bloc’s catch-up strategy, not a property of the technology and not a gift of the market, and it is on loan for as long as that strategy holds.

That should unsettle any government treating advanced AI as something it can buy from allies when the need arises. June showed what buying from allies is worth: the United States cut its own allies’ citizens, and its own employees, off from a model overnight, with a rationale it declined to fully explain. You cannot rely on an old alliance to keep you inside the perimeter, and you cannot rely on the market to hold the floor, because the floor is someone else’s decision to revoke. The only position that survives both failures is the capability you built yourself.

This is a case not for a research programme but for the sustained national investment (compute, talent, academic funding, the entire stack) that a country reserves for things it cannot afford to lose, and it has to begin before the chips land, because capability bought late arrives after the outcome is already decided. The achievable goal for most nations is not winning; it is being undisplaceable, the credible alternative whose existence is the only thing that keeps a global stalemate standing. But you reach that position by aiming past it, at dominance, because aiming at parity lands you underneath it. Treat this as what it is: the industrial revolution and the nuclear arms race arriving as a single event. One half decides whose economy compounds for the next century. The other decides who does the coercing and who absorbs it. Sit this one out and you do not get a smaller seat at the table: you get none, and you learn which it was only once the table is already set.