(Yicai) Oct. 27 -- Models from DeepSeek, one of China's artificial intelligence champions, have topped the call volume ranking on application programming interface aggregation platform OpenRouter for the third week in a row, while Chinese AI models led US rivals for the 26th consecutive week.

DeepSeek's recently released V4.1-Flash-1017 ranked first with call volume at 33.6 trillion tokens in the week ended Oct. 25, followed by V4.1-Flash, according to data OpenRouter released yesterday. MiniMax-M3.1-Flash, Z.AI's GLM-5.3-Flash, and Tencent Hunyuan's Hy4 preview, all Chinese models, rounded out the top five.

Global LLM call volume on OpenRouter rose almost 15 percent to 188 trillion tokens week on week, while also surging from 5.51 trillion tokens a year ago. Among the top nine, Chinese AI models saw call volume jump 18 percent to 93.54 trillion tokens, while that of US counterparts logged a 29 percent increase to nearly 22.5 trillion tokens.

OpenRouter does not represent the entire AI model market, with its data counting only traffic routed through the platform and excluding vendors' own application programming interfaces, cloud platforms, on-premises deployments, consumer-facing apps, and web products. This means the figures cannot be directly equated with global market share.

In addition, a larger share of call volume for Chinese models does not necessarily mean they are generating more revenue, with overseas models still holding a clear lead in OpenRouter's spending share across different tasks. In the overall market structure, general-purpose tasks remain the core segment, accounting for 32 percent of spending, followed by Agent workflows at 31 percent and code-related tasks at 27 percent.

Classification is an important general-purpose task in which models need to identify the category of a given request, such as recognizing customer service intents, tagging documents, or determining which workflow a request should enter. In this task, Anthropic's Claude Opus 5.5 ranked first with a 10.3 percent share, followed by OpenAI's GPT-5.6 Sol and Google's Gemini 3.1 Pro, while V4.1-Flash was the only Chinese model on the list.

Chinese model providers have already used cost-effectiveness to capture traffic entry points and enter more users' workflows, but whether they can turn tokens into revenue and achieve higher profits still depends on actual payment conversion and inference costs. In addition, their models still lag in the ability to reliably complete more difficult tasks. With stronger capabilities, they would also gain greater pricing power.

Editor: Martin Kadiev