Let’s talk about sixty-eight cents.
That is what an American business paid, on average, for a million tokens of AI in August. In March, the same million tokens cost a dollar fifteen. Down roughly 41 percent in six months — and that is not list price, that is Ramp’s index of money that actually moved out of company accounts.
Second number. In early August, 53 percent of the AI usage those companies bought went to frontier models. By September, 45 percent. The share going to the biggest, most expensive, most capable systems is falling.
Third. The top 1 percent of AI spenders — the cohort that carries about 80 percent of OpenAI’s and Anthropic’s enterprise revenue — cut their spend per employee by nearly 10 percent in August.
So nobody stopped using AI. They stopped paying the top price for it.
Now the part that matters if you run something small, because this is where the door opened.
Two years ago, if you wanted capability close to the frontier, you had one path: rent it. You paid per token, you sent your data into somebody else’s building, and your product sat on top of a model that could be repriced, limited or discontinued while you slept. The people who owned the model owned your ceiling.
That path is now optional.
Follow the buyers with the most money, because they are the ones who can afford to do it properly. Thomson Reuters built its own model, Thomson-1, on top of a system they developed by adapting Alibaba’s open-source Qwen. It now handles document review that used to run on Claude. Their CTO said the quiet part out loud: you do not need ever-larger, more expensive models to get useful results. Start from a strong open foundation, specialise it deeply, and you get capable AI at a lower cost.
Harvey is more interesting. Harvey is the legal AI company backed by OpenAI, Sequoia and Andreessen Horowitz, and it spent years building by customising closed models from Anthropic, OpenAI and Google. Its new model, Harvey Tenet, was post-trained on top of an open-weight Kimi K3 — and it outperformed both its base model and the American frontier systems on complex legal agentic work.
Read that again. A company funded by the incumbent got better results by starting somewhere else. That is not a hedge. That is a decision.
And the pricing is not subtle. Z.AI’s GLM-5.3-Flash went out at fifteen cents per million tokens in, fifty cents out. Hugging Face found that in almost every month of 2026, the largest and most capable open model came from a Chinese lab. One venture partner described what he is seeing in his own portfolio plainly: companies building industry-specific foundation models on open source, fine-tuned on their own specific data sets, and not using frontier models for all of the work.
That is the whole story. The frontier is where you go for the hardest 5 percent. Everything else, you buy the weights and keep them.
Now the honest complications, because there are three and they matter.
First: this is early, not won. Only about 3.6 percent of businesses on Ramp’s platform are using open source or Chinese models. The giants still own the market — Anthropic around 43.5 percent share, OpenAI around 39.7 percent. The shift is a direction, not an eviction. Anyone telling you the labs are finished is selling something.
Second: the ceiling on price has been found, and it was found by customers, not by competitors. Anthropic’s most advanced model, Fable 5, is priced at roughly ten dollars per million tokens, about twice OpenAI’s top model — and it accounts for only 6 percent of the tokens businesses buy from Anthropic and 11.4 percent of the dollars. The best model on the market is being treated as a specialist tool, not a default. Ramp’s economist put it plainly: businesses are not willing to pay a premium for the best when a cheaper one is good enough. That is a brutal sentence for anyone whose whole business is being the best.
Third, and this is the trap: open weights are not open source.
Downloading a model is easy now. Understanding how it was made, or changing it at its root, is another matter entirely. You get a binary and a licence. You do not get the data, the recipe, or the freedom to fix what is broken inside it. Meanwhile the neutral noticeboard of the open-model world — Hugging Face, eighteen million developers, three million models — was bought for $12.93 billion in September by the company that sells the chips. The layer that made open models easy to find is now owned by the layer that gets paid when they run.
And underneath all of it the arithmetic is cracking. Morgan Stanley has flagged vulnerability for up to 300 billion dollars in bonds financing the neocloud buildouts — companies that borrowed to build data centres before they had tenants. If the price of a token keeps falling, that debt gets very uncomfortable. The industry told itself that demand for compute was close to infinite and people would pay more for smarter. Instead the token is becoming a commodity, priced like salt. And commodity owners do not get valued at two trillion dollars.
There is also a version of this story that flatters the West too neatly. The reason Chinese open models are winning enterprise attention is not ideology. It is that they are cheap, they can be fine-tuned, and they come with fewer strings on where your data goes. Most South African businesses I know would make that trade in a heartbeat, and they would be right to.
So what is the operator lesson? Three lines.
Your moat was never the model. It was never yours to begin with, and now it is downloadable by a two-person shop in Potchefstroom on a Tuesday afternoon.
What you own is your data, the workflow you have encoded into your business, and the trust your customers hold in you. Those do not show up in a download. That is the part worth spending money on, and it is the only part a pricing war cannot erase.
And take the download — but read the licence before you build a business on it. Know what you can change and what you cannot. Know what happens if the licence changes next year. And know that the moment you cannot leave, you are not a builder any more.
You are a tenant.
The moat did not fall. It was uploaded.