The AI model market has, for the first time, become a real market — and OpenAI and Anthropic are doing everything they can to stop it from staying one.
What the Journal calls “thrift-maxxing” is the plain-language term for what happens when competition returns to a market that had been captured by a handful of premium-priced incumbents. Companies are switching to cheaper models, including Chinese-made ones, because the models work and they cost less. The vendors’ response — partnerships, tens of thousands of dollars in incentives, and heavily subsidized AI usage to lock in customers — is the oldest trick in the monopoly playbook: when the moat starts leaking, you don’t improve the product, you lock the gate.
Cursor’s test tells the story in numbers. Building a web browser from scratch on OpenAI’s GPT-5.5 cost over $10,000. Doing the same with Cursor’s Composer model plus Anthropic’s Opus 4.8 cost $1,339. The cheaper combination was not a compromise; it was a cheaper result for the same task. When Marty Kausas, CEO of the AI customer-support platform Pylon, says “there’s zero loyalty,” he is not describing a market failure; he is describing a market functioning as it should, with buyers exercising choice over a product that is, at this point, a commodity.
The LLMs — all of them — are, at this stage, interchangeable for routine work. The gap between the premium model and the cheap model is a gap in price, not in quality. That is the finding the companies running the premium tier would rather you didn’t notice, and it is the finding that explains why the stock-market valuation of OpenAI and Anthropic depends on your not noticing it. The “bezzle” — Galbraith’s term for the gap between the fraud and its discovery — is not that the models don’t work; it is that the models work about as well as each other for routine tasks, and the premium pricing says otherwise.
The forces — competition, regulation, and self-help/interoperability — are all in play here, and the play is instructive. Competition is the engine: Chinese models like Kimi from Moonshot AI have achieved parity on specific benchmarks, and roughly 50 percent of Hex’s customers adopted Kimi in the last two weeks, per CEO Barry McCardel. The Shanghai conference — where Chinese AI models topped U.S. rivals — was the technical record; this is the business model seeing the consequences of that technical record.
Self-help and interoperability are doing the structural work that antitrust enforcement can only dream of. Telnyx’s David Casem, facing a $100,000-a-day Anthropic bill, switched to a family of models from the Chinese startup Z.AI, using Anthropic’s Fable as a conductor and OpenAI’s Sol for review. The architecture is a patchwork of cheapest-whoever-works, not loyalty to a single vendor. Harvey trained its own model, GLM-5.2, and equipped it with a tool that calls Anthropic’s Fable when a task is hard. Zoom’s CTO, Xuedong Huang, has been fine-tuning Meta’s open-weight Llama model for three years alongside a mix of Anthropic, OpenAI, and open models. The pattern is the same everywhere: the user base is building comcom — competitive compatibility — by stitching together the cheapest possible combination of models that works, and the vendors are scrambling to prevent that stitching from becoming permanent.
The regulatory dimension is where the real danger lies. OpenAI and Anthropic have accused Chinese AI startups — DeepSeek, MiniMax, Moonshot AI — of ripping off their technology. Some Trump administration officials have suggested a ban on such models. The geopolitical framing is the political version of the vendor lock-in: the argument is not that the models are worse, but that they are from the wrong country, and therefore should be banned. The companies that are building the model market are the ones whose models work; the companies whose models don’t cost what the vendors charge for them are the ones whose customers are switching.
The open-weight letter — signed by Nvidia, Microsoft, and Palantir, among others — urging policymakers to exercise caution on restrictions, is the clearest signal yet that the market’s own structural logic is at odds with the vendors’ preferred regulatory outcome. The companies that sell the chips and the infrastructure are not worried about competition; they are worried about the vendor-tier pricing model collapsing.
The factual record is on the side of the customers, and the evidence is plain. Mike Saeks, field CTO at Cursor, said the best model for a task used to change every few months; now it feels like it’s happening multiple times per week. The pace of change is not a threat to the market; it is the market. The threat is to the vendors who built their business on the premise that a premium model would hold its premium for years. That premise was always a bezzle, and the customers are just now figuring it out.
The structural truth is simple: when competition works, prices come down. The companies that benefited from the absence of competition — OpenAI, Anthropic — are the ones now trying to prevent it from working. The companies that never had the monopoly — the Chinese startups, the open-weight projects, the companies stitching together their own patchworks — are the ones making the market function. The question is not whether the AI model market will open; it has. The question is whether the vendors will be allowed to shut it.