Chinese LLMs Lead Global Token Usage, DeepSeek-V4-Flash Surges
According to a report by Pandaily, OpenRouter data indicates Chinese large language models collectively processed over 34.25 trillion weekly tokens, securing the top four positions globally. DeepSeek-V4-Flash, in its official release, advanced to the premier spot, showing a 570 percent increase in token usage from the previous week. This marks a sustained period of Chinese models leading global token metrics.
Artificial intelligence laboratories based in China currently occupy the leading four positions on the global model token usage chart. According to the latest OpenRouter dataset, a recent seven-day period in Aug. recorded global model usage at 69 trillion tokens, representing a 21.48 percent rise from the preceding week. Chinese models contributed 34.25 trillion tokens to this total. Models originating from the United States registered 9.17 trillion tokens during the same period. This marks the fifteenth consecutive week that Chinese models have maintained the leading position in worldwide token count.
And the usage leaderboard provides further detail. DeepSeek-V4-Flash-0731, identified as the official release version, secured the primary slot with 8.83 trillion tokens. This represented a 570 percent increase in usage from the previous week. Tencent Hy3 occupied the second position, registering 8.05 trillion tokens and showing 67 percent growth. DeepSeek-V4-Flash-0423, the earlier preview version, descended to third place, accounting for 5.88 trillion tokens. Xiaomi MiMo-V2.5 subsequently fell to fourth place, with 5.39 trillion tokens. Consequently, all four of the global top positions are presently held by Chinese laboratories.
The DeepSeek official release was made available to the public in late July. The model's fundamental structure and its parameter count remained consistent with the preview version. However, the post training recipe was re executed from its initial stages. This process resulted in notable improvements across both mathematics and coding benchmarks, which contributed to a rapid shift in developer traffic. According to Artificial Analysis, the official V4-Flash model now completes a single task benchmark for an average expenditure of 0.03 US dollars. This cost is approximately one one hundredth of the expense associated with Anthropic's Claude Fable 5 flagship model. It positions the V4-Flash as the most cost effective major model available globally.
Changes within the market rankings were also observed. MiniMax M3, which held seventh place the previous week, and Stepfun Step 3.7 Flash, previously in ninth place, both exited the global top ten during the recent period. OpenAI's GPT-5.6 Luna entered the top five, recording 4.43 trillion tokens and demonstrating 128 percent growth. This occurred after the organisation reduced its price by 80 percent for enterprise applications.
So, the overall assessment indicates that Chinese open source models have transitioned their competitive edge from primarily price based to usage based. A consistent fifteen week period of global leadership, the occupation of all four top positions, and a flagship model costing one one hundredth of its major competitor collectively point to the same conclusion. The open source Chinese artificial intelligence stack is no longer competing solely on price. It demonstrates success across a combination of price, processing capacity, and benchmark quality.
Chinese large language models have led global token usage for fifteen consecutive weeks.
DeepSeek-V4-Flash-0731, the official release, achieved the top position in global token usage, showing a 570 percent increase.
Chinese laboratories now hold the top four positions on the global model token usage chart.
The DeepSeek V4-Flash model is noted for its low cost, approximately 0.03 US dollars per single task benchmark.
OpenAI's GPT-5.6 Luna and Google's Gemini 3.6 Flash also saw significant increases in usage after price adjustments and market entry.
Source: Pandaily
