Chinese Models Reach 46% of Token Usage at US Companies — and the Reason Is Price
Chinese models' share of token usage at US companies via OpenRouter has topped 30% every week since February, peaking at 46%, versus 11% a year ago. The reason: a price gap of up to 55x. A balanced reading.
One number sums up a shift that has rattled the AI industry's assumptions: since February 8, 2026, the share of Chinese AI models in token consumption by US companies via the OpenRouter platform has not fallen below 30% in any week, peaking at 46%. For comparison, the average over the prior twelve months was just 11%, dropping to 4.5% in the first half of 2025. The driver is not politics, but price: open models that rival the frontier at a far lower cost.
What Is OpenRouter, and Why Does This Number Matter?
OpenRouter is a routing platform that lets developers pass their requests to any of more than 400 models across more than 60 providers, by simply swapping the access endpoint and key. This makes it a sensitive mirror of what developers actually choose when switching carries negligible cost. And because the popular Chinese models are "open-weight," there is no commercial "lock-in": you download them or access them via any platform, lowering the switching cost to nearly zero.
The Whole Story Is in the Price Gap
The core of the shift is purely mathematical. OpenRouter data says open Chinese models are 60% to 90% cheaper than the leading equivalents from OpenAI and Anthropic. The numbers are striking at the extreme: DeepSeek's V4 Flash model at about $0.14 per million input tokens, versus $5 for GPT-5.5 and Claude Opus 4.8 — a gap of roughly 55 times. And even in the flagship tier, Zhipu's GLM-5.2 runs at about $1.40 input and $4.40 output, three to six times cheaper than Opus. As one official summed it up: "Price is doing the work here; when a task doesn't need the best model, teams route it to the cheapest one that's good enough."
Not an Experiment, but a Move to Production
What reinforces the seriousness of the shift is that it has left the experimentation phase. The startup Lindy moved 100% of its traffic from Claude to DeepSeek, saying it will save millions. And GLM-5.2, launched in June, saw the fastest adoption of any model on the Vercel platform in 2026: its daily token volume grew about 27x and its customer count about 80x in its first week. According to Financial Times reports, Chinese models now process about 18 trillion tokens a week across the platform's top nine models, versus about 5.5 trillion for US models — after the US led in January.
Why Now? Pressure From Two Directions
The shift comes from a double pressure. On one hand, US model prices are rising as usage explodes; Uber, according to reports, exhausted its annual AI budget just four months into 2026. On the other hand, users have shifted from a "maximize usage at any cost" mindset to a return-on-investment mindset. And the regulatory context adds an ironic dimension: as the US tightens access to its most powerful domestic models (like the temporary suspension of Fable 5 and Mythos 5), the market has been unintentionally pushed toward cheaper external alternatives available without restrictions, since Chinese labs face no constraints on distributing their open models globally.
A Balanced Reading: Necessary Caveats
Despite the data's importance, it should be read carefully, as the sources themselves cautioned. First, OpenRouter is one platform, not the whole market, and the reports do not clearly separate "enterprise production traffic" from exploratory "developer experimentation." Second, the higher figure (46%, or 61% in a particular week) measures the most-used models on the platform, not its full catalog. Third, Claude and ChatGPT still dominate on other platforms, especially in regulated sectors. The precise conclusion: a real and growing slice of US demand is now behaving "like a commodity" routed to the cheapest good-enough option, but this does not yet mean the entire "market standard."
The practical lesson of this shift is broader than brand competition. The infrastructure that wins is the "routing and observability" layers, and the winner is every buyer whose cost model was written when cheap open weights did not yet exist. But relying on any single provider — Chinese or American — carries its own risks, as the Fable 5 suspension showed the fragility of building on a single model. The more mature advice for teams: measure end-to-end product metrics (latency, safety, hallucination rate) rather than leaderboard position alone, and balance cost savings against governance and vendor-evaluation overhead.