DeepSeek jumps 801% as Chinese AI models route 4x the US on OpenRouter

What's the deal? Chinese AI models processed 93.54 trillion tokens on OpenRouter during the week of October 19–25 — roughly 4.2 times the 22.47 trillion routed through US models, according to OpenRouter data cited by Chinese outlet Caixin. DeepSeek's updated V4.1-Flash-1017 topped the global chart with 33.6 trillion tokens, an 801% jump from the prior week.

Why now? The surge followed DeepSeek's October 17 model update, built on the V4.1 Flash released September 10, which aims to improve reasoning, coding, automation, and vision tasks. Rivals moved in parallel: MiniMax released its M3.1 Flash on October 19, claiming nearly tripled inference speed and 70% lower costs.

Who's winning? Chinese models swept all five top spots. DeepSeek's V4.1 Flash came second at 18 trillion tokens, MiniMax's M3.1 Flash third at 13.7 trillion, Zhipu AI's GLM-5.3 Flash fourth at 13.2 trillion, and Tencent's Hy4 Preview fifth at 8.97 trillion.

By the numbers: Global usage hit 188 trillion tokens over the period, up 14.63% week on week. Chinese models accounted for 49.8% of the total; US models held 11.9%. It marks the 26th straight week Chinese models have led on the platform.

What could go wrong? Routed token volume is not market share or revenue. OpenRouter captures only traffic sent through its service — models accessed directly are excluded — and reporting indicates US frontier providers still attract more spending.

The signal: Competitive pricing and strong performance on coding and agentic workloads are pulling developers toward Chinese open-weight models, a trend also reported by CNBC across OpenRouter and Vercel. The lead is concentrated in a fast-moving, price-driven corner of the market — but it is widening.

Read more: ajupress.com, globaltimes.cn, CNBC