News · 2026-06-25
Are closed AI models overpriced luxury goods?
Here is a question the AI industry would rather you didn't dwell on: if you can download a freely available model that does most of what the expensive subscription model does, why does the expensive one cost so much more? An essay by James O'Claire, The Unbearable Cheapness of Open Weight Models, takes that question seriously and arrives at an uncomfortable answer.
First, some plumbing. "Open-weight" models are ones whose finished brains are published for anyone to download and run, as opposed to "closed" models you can only rent through a company's service. Our open-weight models explainer covers the distinction in full. O'Claire's starting observation is that the price gap between the two has become enormous. By his accounting, the leading openly available models, several of them from Chinese labs, charge a tiny fraction of what the big Western labs charge for a comparable amount of work. We won't fixate on the exact multiple, but the claim is that it's not a little cheaper; it's dramatically cheaper, the kind of gap that demands an explanation.
His explanation is that the high prices aren't really about cost; they're about positioning. He argues the leading closed labs are, in effect, selling a luxury product, manufacturing scarcity and leaning on premium branding rather than competing on price, the way a designer handbag costs many times what a sturdy unbranded one does despite carrying the same things. If that's right, the price of a frontier API reflects a moat the companies want to protect, not the raw cost of running the model.
Then comes the sharper, more political claim. O'Claire worries that the Western labs have found a convenient lever to protect that moat: fear of China. If openly available Chinese models are the thing undercutting your prices, then framing those models as a security threat, and pushing the government to restrict them, conveniently removes your cheapest competition while wrapping the move in the flag. He ties this directly to the running accusation that Chinese labs have been "distilling" Western models, training on their outputs to copy their abilities, an accusation that has surfaced repeatedly, including earlier reporting from TechCrunch on Western labs raising exactly these alarms. His point isn't that distillation is fine; it's that "protect our intellectual property" and "protect our prices" can be the same incentive wearing two different hats.
His constructive ask is for "true" open source, not just published weights but open training data too, so the whole recipe is inspectable, and he points to academic and government-backed efforts as examples of what that could look like.
Why it matters: this is the economic and political frame underneath one of 2026's defining tensions, a cheap, open commodity floor pressing up against an expensive, closed premium, now spilling into Washington. It reframes the distillation fight: what gets described as a clean story about intellectual-property theft is also, unavoidably, a story about who gets to keep charging a premium. The same dynamic shows up in our earlier coverage of how open weights became an insurance policy for companies wary of depending on a single vendor.
The honest caveat is that this is an opinion piece, and it should be read as an argument, not a verdict. The eye-popping price gap mixes together very different things, reliability, support, safety guarantees, and the real cost of running a model at scale, that a pure per-word comparison flattens. A closed model's price isn't only branding. But the essay is a useful corrective to taking either the "premium models are simply worth it" or the "open models are a national security threat" story at face value. Both, it suggests, deserve a harder look at who benefits.