Chinese AI models hit 57–67% of OpenRouter tokens — up fivefold in seven months
Chinese AI models handled 57 to 67 percent of the tokens processed on OpenRouter, the largest service for routing requests across AI models, in the week of September 14, 2026, up from 6 to 13 percent in February, CNBC reported on September 26. On Vercel, a web development and hosting platform, their share climbed from 11 percent in January to 55 percent in August. Across the 82 countries OpenRouter groups as the Global South it was 67 percent. Staff at both platforms said open models from DeepSeek, Z.ai and Alibaba became capable enough for multi-step agent work such as coding, and from that point price decided. American models still take more of the money spent, and two House committees are investigating
The three lines
- Numbers — OpenRouter token share from 6–13% in February to 57–67% in mid-September; Vercel from 11% to 55%; Global South 67%
- Why — once Chinese open models could handle coding and agent tasks, the price gap became the deciding factor, both platforms say
- Caveat — U.S. models still get more total spending and harder tasks; two House committees are investigating
Key questions
- What share of AI usage do Chinese models have
- **More than half of tokens on two major developer platforms.** | Platform | Earlier | Latest | |---|---|---| | OpenRouter (all) | February **6–13%** | week of September 14 **57–67%** | | OpenRouter (Global South, 82 countries) | — | **67%** | | Vercel | January **11%** | August **55%** | The OpenRouter figure appears as a range, likely because of how model providers are classified by nationality. Vercel did not give a regional breakdown.
- Why are developers switching to Chinese AI models
- **Once quality crossed a usable threshold, price took over.** | Factor | Detail | |---|---| | Performance | DeepSeek, Z.ai and Alibaba models improved sharply at coding (CNBC) | | Workload | OpenRouter's Peter Walker: capable of advanced multi-step agent work, especially coding | | Price | Vercel's Harpreet Arora: past the quality bar, the price gap is a strong deciding factor | | Open weights | many hosts can serve the same model and compete on price | U.S. models still lead most benchmarks, but the gap is narrowing.
- Do Chinese AI models lead in spending too
- **No. They lead in volume, not in money.** | Measure | Ahead | |---|---| | Tokens processed | Chinese models | | Total spending | U.S. models | | Complex tasks | more often U.S. models | The market is splitting: cheap, high-volume work goes to Chinese open models; costly, error-sensitive work stays with U.S. models. U.S. companies account for about half of all OpenRouter tokens.
The most-used AI models in the developer market changed nationality in seven months. According to platform data reported by CNBC on September 26, 2026, Chinese models handled 57 to 67 percent of the tokens processed on OpenRouter — a service that routes requests to hundreds of AI models through one account — in the week of September 14. In February the figure was 6 to 13 percent.
1. The numbers — two platforms, one direction
| Platform | What it is | Earlier | Latest |
|---|---|---|---|
| OpenRouter | router for 400+ models | February 6–13% | mid-September 57–67% |
| OpenRouter Global South | firms in 82 countries in Latin America, Africa, Asia | — | 67% |
| Vercel | web app development and hosting | January 11% | August 55% |
Tokens are the word fragments AI models read and write; token share is a measure of how much work was actually handed to each model.
The two platforms serve different users. OpenRouter is where developers switch between models; Vercel is where they build and ship web services. Two different places moving the same way at the same time is what gives the numbers weight. U.S. companies account for about half of all OpenRouter tokens.
2. Why — past the quality bar, price wins
| Explanation | Who |
|---|---|
| DeepSeek, Z.ai and Alibaba sharply improved coding performance | CNBC |
| Chinese models now handle advanced agent-type use | Peter Walker, OpenRouter |
| Once quality thresholds are met, the price gap decides | Harpreet Arora, Vercel |
The key is a threshold effect. When a cheaper model falls short, developers pay for the U.S. one. Once a Chinese model can also edit files and run tests as a coding agent, developers look at the bill. Agents burn hundreds of thousands to millions of tokens per task, so unit price becomes total cost.
| Model (per million tokens) | Input | Output |
|---|---|---|
| Claude Opus 5.5 | $4 | $20 |
| GPT-6 Sol | $2 | $10 |
| GPT-6 Luna (light) | $0.10 | $0.50 |
U.S. labs cut prices sharply on September 22, and light U.S. models now start at 10 cents per million input tokens. So the comparison developers make is not unit price alone but how hard a task a model can handle for the money. By the accounts of Walker and Arora, Chinese open models won share by doing coding-agent work while costing far less than top U.S. models. CNBC did not publish model-by-model price comparisons. Because most Chinese models release their weights, many hosts serve them and compete on price.
3. What worries Washington — and where the U.S. still leads
| Issue | Detail |
|---|---|
| House inquiry | two House committees are investigating the spread of Chinese AI |
| Export-control workarounds | concern about chips accessed through overseas data centers |
| Distillation | suspicion that U.S. model outputs are used to train rivals |
| Spending | U.S. models still draw more total spending |
| Task difficulty | complex work more often goes to U.S. models |
The market is splitting into volume for China, money for the U.S. High-volume, simple-to-medium work goes to cheap Chinese models; work where mistakes are costly stays with expensive U.S. ones.
The 67 percent in the Global South tells another story: firms most sensitive to dollar prices moved first. Washington's effort to build an AI bloc around U.S. chips and models is running into usage that follows price. The same week, the U.S. and Chinese presidents agreed to a dialogue on AI risk but left chip export controls in place.
4. What remains open
- The range — why OpenRouter's figure spans 57–67% could not be confirmed from primary data.
- The inquiry — which two committees and what they are examining are unconfirmed.
- Korea — no public data shows Korean developers' use of Chinese models.
- Data location — running open weights on your own servers keeps data at home; calling a Chinese company's API does not. How the numbers split between the two is not disclosed.
Sources
- CNBC — Chinese AI models surge in global popularity, and Washington is worried
- Digital Today — China's value AI models speed global spread, raising tensions in Washington
- Superpower Daily — Chinese AI models gain majority token share on two platforms as House panels investigate
- Trending Topics — Open-weight models from China are capturing a growing share of AI usage