The Frontier Can't Keep What It Sells

Open weights are a season behind and fifty times cheaper. That’s a problem the best models can’t out-build, and the reason the doom case doesn’t end where you’d expect.

Every company paying for a frontier model is making the same wager: that the best model is worth a premium. It’s a reasonable wager right up until you notice how strange a thing “the best” is to pay for. You don’t pay extra for the best hammer once the cheaper one drives the nail. Capability has a ceiling defined not by the model but by the task, and for most tasks that ceiling arrived a while ago.

Here’s the part of that instinct the numbers back up. Epoch AI tracks how far open-weight models trail the closed frontier, and as of mid-2026 the gap is about four months, and it’s asymmetric. On hard reasoning, the closed labs still hold a real lead. On coding and agentic work, the gap has effectively closed. You can run an open model and not feel the difference on most of what you ship. Open weights aren’t the budget option anymore, they’re the volume leaders. Chinese open-weight providers now push close to half the tokens flowing through aggregators like OpenRouter, a share that sat near zero a year ago. And they do it for somewhere between a thirtieth and a fiftieth of the frontier price.

So the question is the right one: when the model already does everything you need, what exactly are you buying with the premium? A number you can’t feel?

But the case is harder than “open will catch up,” and harder in a way that should worry the labs more than any benchmark could.

You can’t sell the teacher without making the student

There’s a technique called distillation. You take a strong model, the teacher, feed it a large volume of carefully built prompts, capture its outputs, and train a smaller, cheaper model, the student, on those outputs until the student behaves like the teacher at a fraction of the cost. The crucial detail, in one lawyer’s framing of the recent disputes, is that it turns a public inference service into a training corpus for a rival. You don’t need the weights. You don’t need to break anything. Selling API access is enough to hand over the training signal.

This isn’t theoretical. In February, Anthropic accused three Chinese labs, DeepSeek, Moonshot, and MiniMax, of industrial-scale distillation: roughly 16 million exchanges across some 24,000 accounts, aimed at reasoning, agentic tool use, and coding. OpenAI leveled similar accusations at DeepSeek a year earlier. The companies named dispute the characterization, and you can argue about any single case. What you can’t argue is the economics underneath. Berkeley researchers recreated a reasoning model for $450 in a day. A Stanford and UW group did a version for under fifty dollars. Databricks’ CEO put it plainly: the technique is extremely powerful, extremely cheap, and available to anyone.

Sit with what that means for the business. A frontier lab’s product is access to the teacher. The act of selling that access is the act of distributing the training signal for the teacher’s replacement. You are, structurally, in the business of seeding your own competition, and the better your model, the more valuable the thing you hand out with every call.

The labs know it, and the tell is in how they’ve responded. OpenAI has said it now runs “a careful process for which frontier capabilities to include in released models.” That sentence is doing a lot of quiet work: the leading lab is deliberately holding some of its best work back from the product to keep it from leaking. When the defense becomes “don’t ship the full capability,” the product and the moat are already in tension.

The prices weren’t the whole story

Set distillation aside for a second. The prices you’re comparing aren’t the full cost picture.

Frontier inference is being sold somewhere between ten and twenty times below what it costs to serve, propped up by venture capital and hyperscaler cross-subsidy. The labs whose numbers went public this year were spending two to three times their revenue; compute, not salaries, is the dominant line. By several accounts the largest player loses money on every dollar of revenue it books. None of this is a scandal. It’s a land grab, the oldest playbook in software. But land grabs end.

And here’s the cruel geometry: as the subsidy unwinds, frontier prices are going up. The race to zero is happening on the open-weight side. The closed labs are adding capability and raising prices at the same time. Enterprises are already feeling it: companies burning through annual AI budgets in a single quarter, capping per-employee spend, standing up dashboards to meter token usage like electricity. The widely shared forecast is another 30 to 50 percent on frontier API prices inside two years as the economics normalize.

So the premium you’re choosing not to feel today gets more expensive at exactly the moment the free alternative gets good enough to make the comparison awkward. That’s not a gap. That’s a vise.

Where the doom case gets too clean

If I stopped here I’d be writing the same triumphant “open wins, pay nothing” post that’s all over the timeline, and I don’t believe that one, because two things complicate it badly.

First: the gap isn’t closing. It’s holding. Four months behind in mid-2026, slightly wider than the three-month average of the prior two years. “Open catches up” is the wrong tense. The accurate version is “open stays roughly a season behind, permanently, because the frontier keeps moving.” And whether a season behind is fine depends entirely on the job. For drafting an email, last season’s model is fine forever. For an agent acting unsupervised against a production system, the last few points of reliability are the difference between something that needs a human watching it and something you can trust to run, and that difference is worth a fortune the benchmark can’t show you, because value in that regime is non-linear. Eight points of capability can be the whole ballgame.

Second: the labs aren’t really selling models anymore. They’re racing to sell the layer above the model: agents, memory, tool use, integrations, the reliability and the product and the distribution. The base model is becoming a component. Watch the market sort itself out: even the open champions are bifurcating, splitting off closed premium tiers while keeping a free base a generation behind. Everyone is converging on the same shape, a paid frontier sitting on top of a commoditized floor.

Which is why “frontier models are doomed” is the wrong sentence. The model commoditizes; that part is happening and won’t stop. What’s under pressure is the specific bet: spend a fortune to train the smartest model and rent access to it as the product, full stop. Linux didn’t kill software. It moved the money up the stack and sideways into service, and the companies that read the shift early did fine. The frontier labs are trying to make that same move, from “we sell the smartest model” to “we sell the system you build on it,” before their core asset becomes free. Some will manage it. The pure token-vendor play is the one that looks hardest to defend.

The part that keeps me up

Here’s the knot I can’t untie, and I’d rather leave it honest than pretend I’ve solved it.

The open-weight ecosystem doesn’t replace the frontier. It feeds on it. Distillation needs a teacher. A model that’s four months behind is four months behind something. The cheap, good-enough, fifty-times-cheaper models that make the doom case so persuasive are, every one of them, chasing a frontier that someone else paid to discover.

So follow the doom case all the way down. If the frontier can’t be monetized, because the moment it’s sold it’s copied, and because its unsubsidized price is one most buyers won’t pay when good enough costs a fiftieth as much, then who funds the next ten-figure training run? The ending isn’t “the labs lose and we all get cheap intelligence forever.” The ending is that the engine everyone’s been distilling runs out of fuel, and the field settles into a permanent plateau a season behind a frontier that stopped advancing. The open models would be mortgaging a future they don’t generate.

I don’t know how that resolves, and I’m suspicious of anyone who says they do. What I know is that I route my own work the way most builders do now: open weights for the bulk of it, the frontier held back for the handful of jobs where the gap is the point. Which puts me on both sides of my own thesis: the person proving it, because I won’t pay the premium for most of what I do, and the person it would strand, because I still need a frontier to exist for the work that matters. Those two facts don’t reconcile. I’ve stopped expecting them to.