DeepSeek Tops OpenRouter Call Volume Ranking for Three Consecutive Weeks, but High-Value Task Spending Still Dominated by Overseas Models

On October 26, the large-model API aggregation platform OpenRouter published data on the call volume on its platform last week. On the call-volume leaderboard, the model DeepSeek updated not long ago debuted in first place, and DeepSeek's models have now ranked first on the platform's weekly list for three consecutive weeks. In addition, other vendors near the top of the list include MiniMax, Zhipu, and Tencent, with Chinese models sweeping the top five.

Specifically, in the week from October 19 to 25, the total global call volume for large AI models was 188 trillion tokens, up 14.6% week-on-week. Among the top nine, the weekly call volume of Chinese large AI models reached 93.54 trillion tokens, up 18.1% week-on-week; over the same period, the weekly call volume of US large AI models was nearly 22.5 trillion tokens, up 29.4% week-on-week.

This marks the 26th consecutive week that Chinese large models' weekly call volume on OpenRouter has exceeded that of the United States, ranking first globally. In the top-nine call-volume list, Chinese models occupy six places, while overseas models hold three.

Specifically, last week, DeepSeek's V4.1-Flash-1017 model, updated on October 17, reached a weekly call volume of 33.6 trillion tokens, up 801% week-on-week, ranking first in call volume; second place went to DeepSeek's previous-generation model V4.1 Flash, with a weekly call volume of 18 trillion tokens.

MiniMax's new model M3.1 Flash, released on the evening of October 19, debuted in third place. Zhipu's GLM 5.3 Flash and Tencent Hunyuan's Hy4 preview ranked fourth and fifth, respectively.

Looking at the overseas model data, OpenAI's GPT-6 Luna ranked sixth with a call volume of 8.35 trillion tokens, and Anthropic's Claude Opus 5.5 had a call volume of 8.27 trillion tokens, ranking seventh.

It should be noted that OpenRouter does not represent the entire global large-model market call volume. Its data only counts traffic routed through the platform; vendors' own APIs, cloud platforms, private deployments, and applications and web products directly facing consumers are not included. Therefore, these data cannot be directly equated with global market share.

However, as an openly accessible model API aggregation platform, OpenRouter's data can serve as a reference for observing market trends. The official page notes that the above weekly call-volume leaderboard data shows the sum of each model's prompt and completion tokens, including reasoning tokens.

Over a longer time span, the volume of tokens OpenRouter processes each week grew from 5.51 trillion in late October 2025 to 188 trillion in late October 2026, a year-on-year increase of nearly 33 times. Along with the overall expansion of the market, traffic is also continuing to migrate toward Chinese models.

In the week of October 27, 2025, the model ranked first in OpenRouter call volume was Grok from Musk's xAI, and the list was almost entirely overseas vendors, but now the trend has completely reversed.

English text is being prepared.

For high-value, high-risk, or high-quality-demand tasks, enterprises are still willing to use the world's top models. But in real work environments, there are a large number of relatively low-value and low-risk tasks that are price-sensitive, and this is precisely the market for which Chinese models are better suited.

However, Chinese models gaining a larger share of calls does not mean they are earning more money.

If we look at another leaderboard on OpenRouter—the share of spending by models on different tasks—overseas models are still absolutely ahead.

In the figure above, different colors represent different categories of tasks, and the size of the area represents the amount of money spent. In the overall market landscape, general tasks are the base, accounting for 32.1%, followed by Agent (agentic workflows) at 30.9% and Code at 26.7%.

Classification is one of the important general tasks. In this task, a model needs to identify the category of a call request, for example, identifying the demands in a customer service scenario, tagging documents, or determining which workflow a request should enter.

In this single task, Anthropic's Claude Opus 5.5 ranks first, holding a 10.3% share, followed by OpenAI's GPT-5.6 Sol and Google's Gemini 3.1 Pro in second and third place. The only Chinese model on the list is DeepSeek's V4.1 Flash.

The top-ranked models are basically the most advanced models from each overseas company, which shows that when doing tasks such as classification that require relatively high accuracy, developers spend the most money on top overseas models.

Overall, Chinese model vendors have already used cost-effectiveness to win the traffic entry point and enter more users' workflows, but whether they can turn tokens into revenue and obtain higher profits still depends on specifics such as payment and inference costs. The ability of models to stably complete highly difficult tasks is also something Chinese vendors need to catch up on; with better capabilities, they will also have greater pricing power.