Here is a number that should stop every U.S. AI lab cold. The share of tokens American companies route to Chinese open-weight models through OpenRouter — a marketplace developers use to shop across dozens of models — has sat above 30% every week since February 8, and has spiked as high as 46%, according to CNBC. Over the prior twelve months, that figure averaged just 11%. In other words, roughly one in three tokens U.S. developers buy on the open market is now going to a model built in China, and on some weeks it is closing in on one in two.
The cause isn’t politics, and it isn’t benchmarks. It is the bill. Open-source Chinese models can run 60% to 90% cheaper than the flagship offerings from OpenAI and Anthropic, and the buyers have noticed. “Price is doing the work here,” Vercel told CNBC. “When a task doesn’t need the best model, teams are beginning to route it to the cheapest one that’s good enough, and the recent wave of models coming out of China is winning that trade.” More than a trillion tokens now flow through Vercel’s own AI gateway every day, and the company says customers have stopped picking one lab and building everything on top of it.
The examples are getting concrete. AI startup Lindy moved 100% of its traffic off Anthropic’s Claude and onto DeepSeek, and told CNBC the switch will save it millions. Z.ai’s GLM-5.2, released to fanfare in June, posted the fastest adoption of any model Vercel tracked all year: daily token volume up about 27x and customer count up about 80x in its first full week. This is the same dynamic that put $800 million behind Together AI to host open models, and the same one DeepSeek is now chasing into its own silicon.
Our take: The frontier stopped being scored on who is smartest and started being scored on who is smartest per dollar. That is the exact metric Google is holding Gemini 3.5 Pro in preview to win, and it is why the U.S. price war is already underway: Anthropic cut Claude Sonnet 5 to an introductory $2 per million input tokens and $10 output through August 31 — undercutting its own Opus 4.8 — and OpenAI has reportedly weighed steep cuts of its own. The uncomfortable truth for the incumbents is that “good enough and 80% cheaper” wins the vast middle of the workload, and the middle is where the volume lives. Being the best model is no longer the same business as being the default one. When even premium tiers start billing by the token, the customer’s next question is always the same: do I need this one, or will the cheap one do?
There is a real ceiling on the trade. Data-security and compliance worries still keep the largest regulated enterprises — banks, healthcare, government contractors — from routing sensitive work to Chinese models, no matter the price. That is the moat U.S. labs have left, and it is narrower than a comfortable one. For everyone else — the startups, the agent builders, the teams running millions of low-stakes calls a day — a list price near $1.40 per million input tokens beats one near $5, and the spreadsheet doesn’t care where the weights were trained.
What to watch
- The weekly share. Above 30% since February is a trend, not a blip. If it holds north of 40% into the back half of the year, “route to the cheapest good-enough model” stops being a tactic and becomes the default architecture.
- The price responses. Anthropic’s intro rates expire August 31; watch whether they hold, and whether OpenAI turns its rumored cuts into real ones. The direction of U.S. pricing is now set in Hangzhou and Beijing.
- The compliance gate. The one thing keeping big regulated buyers on U.S. models is security review, not capability. If Chinese labs clear that bar — or a self-hosting layer clears it for them — the last wall comes down.