News
Chinese AI Models Now Handle Up To 46 Percent Of US Companies' OpenRouter Queries
Rising prices from OpenAI and Anthropic are pushing US companies toward cheaper Chinese open source models. OpenRouter data shows Chinese models' share of queries from American customers has reached as much as 46 percent this year, up from 11 percent at the start of the year.
American companies are increasingly shifting their compute traffic away from OpenAI and Anthropic models toward cheaper Chinese open source alternatives. According to data from the OpenRouter platform, the share of tokens processed by Chinese models in queries from US customers has reached as much as 46 percent this year, up from just 11 percent at the start of the year.
Price is the main driver behind the shift. Chinese models such as Qwen, DeepSeek, Kimi, GLM and MiniMax offer performance close to leading US systems at significantly lower token processing costs. Engineers responsible for infrastructure budgets are increasingly routing production workloads to wherever the economics work out best.
The Cost Math
GLM 5.2, made by China's Zhipu AI, costs about $0.91 per million input tokens and $2.86 per million output tokens on OpenRouter. By comparison, Anthropic's Claude Opus 4.8 is priced at roughly $5 per million input tokens and $25 per million output tokens, meaning the Chinese model is roughly five times cheaper.
Despite that large price gap, benchmark results remain surprisingly close. On the FrontierSWE test, which measures a model's ability to handle long, multi-step coding tasks, GLM 5.2 landed within one percentage point of Opus 4.8. On the Artificial Analysis Intelligence Index, GLM 5.2 ranks first among open models, ahead of Nemotron 3 Ultra, MiniMax M3, DeepSeek V4 Pro and Kimi K2.6.
The Scale Of The Shift
Individual company examples show how fast the migration is moving. Startup Lindy shifted all of its traffic from Claude to DeepSeek, a move it says will save millions of dollars. On the Vercel platform, daily token volume processed by the GLM-5.2 model grew twenty-sevenfold in the first week after it became available.
Coding now accounts for more than half of the total token volume processed on OpenRouter, up from 11 percent at the start of 2025. It's in agentic and coding tasks that the cost difference matters most, since AI agents can burn through as many as 20 million tokens on a single complex coding task.
What It Means For Companies
For Polish companies using language models in production, this represents a real pricing alternative, particularly for large-scale workloads where token costs directly affect product profitability. At the same time, shifting data and queries to models hosted or developed in China raises questions about data security, compliance with EU regulations, and dependence on suppliers outside the West.
OpenAI and Anthropic have not yet announced price cuts in response to the growing competitive pressure, betting instead on quality advantages and integration within their own tool ecosystems. How long US labs can hold current pricing without losing further market share remains an open question.
Sources: CNBC (cnbc.com), Trending Topics (trendingtopics.eu), Technology.org (technology.org)
