Chinese AI models dominate developer platforms, but the picture is skewed

Chinese AI models are taking up a large share of usage among developers who rely on routing services such as OpenRouter and Vercel. But the numbers are hard to extrapolate to the entire AI market.

More than half of the traffic

OpenRouter is a service that gives developers access to many different AI models in one place. According to CNBC, as relayed by the analytics firm Digital Applied, Chinese models made up 57-67 percent of OpenRouter's tokens in the week beginning September 14, 2026. In February, the share was 6-13 percent. A token is a small chunk of text that an AI model reads or writes.

Another tally puts the figure at around 53 percent of OpenRouter's public token volume on September 13. On the Vercel platform, the company says Chinese models' share of traffic reached 55 percent in August, up from 11 percent in January.

According to Axios, as relayed by Let's Data Science, Chinese open models held the top five spots in OpenRouter's weekly token usage. Among them were Kimi K3 and models from DeepSeek, Tencent, Xiaomi, MiniMax and Z.ai.

Kimi K3 in the spotlight

Moonshot AI announced Kimi K3 on July 16, 2026. The model has 2.8 trillion parameters, can work with text and images, and has a context window of 1 million tokens, meaning how much text it can handle at once. Moonshot calls it the world's largest open-weight model, and the weights were scheduled for release on July 27. The price is $0.30 per million cached input tokens and $15 per million output tokens.

According to Moonshot's own figures, K3 scores 77.8 on Program Bench and 93.5 on GPQA-Diamond, and it ranks above GPT-5.6 Sol and Fable 5 on Arena.ai's blind coding leaderboard. The figures are the company's own and have not been independently verified. According to the coverage, early blind tests showed that developers preferred K3 for front-end coding over leading American systems.

Why the numbers should be read with caution

Digital Applied points out that tokens measure the volume of text processed, not revenue, customer counts or quality. The figures only cover traffic through routing services, not direct customers, self-hosted setups or enterprise deployments. Nor should you average across platforms and time periods, since that produces an invented combined number.

The authors Mary Clare McMahon and Konstantin Pilz point to data from the spend management platform Ramp: At least 77 percent of the customers who spent money on AI bought products from OpenAI or Anthropic. Only 6 percent used platforms such as OpenRouter, and 96 percent of them also paid for Anthropic or OpenAI. According to the authors, closed models therefore likely dominate businesses' use of AI. Consumer use is less clear. They also argue that tokens are not all worth the same, and that token counts should not be compared without accounting for the size and capabilities of the models.