Chinese labs now lead developer traffic on OpenRouter, the multi-model API gateway that routes calls from apps and coding agents to models from many providers. Rankings published with usage data through Oct 18, 2026 show Chinese authors leading both request share and weekly token volume. That is a usage measure on one marketplace, not a claim about the full AI market. OpenRouter says so in its own methodology notes.
What OpenRouter reports for the latest week
On the market-share chart for the week beginning Oct 12, 2026, DeepSeek held 24.4% of text requests, ahead of Google at 19.7% and OpenAI at 18.5%. Z.ai (Zhipu) sat at 12.2%, Alibaba's Qwen at 5.3%, Anthropic at 4.7%, and Xiaomi at 2.4%. Meta Llama and Mistral AI each had about 2%. The rest of the authors shared 8.6%.
Those named Chinese authors alone (DeepSeek, Z.ai, Qwen, Xiaomi) add to about 44% of requests that week. Google, OpenAI and Anthropic together add to about 43%. The gap is narrower than the prior week that Korean outlets summarized, but DeepSeek remains the single largest author by request share.
Token volume: the top five weekly models are Chinese
OpenRouter also ranks models by tokens processed (prompt plus completion). For the trailing weekly window, the five highest models are all from Chinese labs:
- DeepSeek V4.1 Flash: 32.8T tokens
- GLM 5.3 Flash (Z.ai): 12.2T tokens
- Step 5 Preview (StepFun): 11.5T tokens
- MiMo-V2.6-Flash (Xiaomi): 9.94T tokens
- Hy4 preview (Tencent): 7.55T tokens
The next two are US products: GPT-6 Luna (OpenAI) at 6.21T and Claude Opus 5.5 (Anthropic) at 6.18T. A second Z.ai model, GLM 5.3, ranks eighth at 5.19T. A second DeepSeek Flash variant ranks tenth.
Tokens and requests are different meters. Token totals rise when a model is chatty or when agents run long loops. Request share counts how often an author's models are called. Both charts point the same way on OpenRouter as of Oct 18: Chinese authors lead, mostly with Flash-class models.
The Oct 5 to 11 week that outlets called 48%
The Dong-A Ilbo, writing on Oct 19, reported that Chinese AI models accounted for about 48% of OpenRouter developer requests from Oct 5 to 11, above the combined 40.3% share of OpenAI, Google and Anthropic. It listed DeepSeek at 24.5%, OpenAI at 18.6%, Google at 17.8%, Zhipu AI at 13.4%, and Qwen at 5.4%. It also said Zhipu's requests rose 107% week over week, and that Chinese companies took all five top spots when models were ranked by data volume that week.
Businesskorea published the same Oct 5 to 11 percentages on the same day. Those figures are secondhand summaries of OpenRouter data for an earlier week. The live rankings page now shows the week beginning Oct 12, with the author shares listed above. Use Dong-A and Businesskorea for the prior-week framing; use openrouter.ai/rankings for the current week.
The Oct 18 spike after DeepSeek's Oct 17 update
On the most recent complete day in the rankings (Oct 18, counted in UTC), DeepSeek V4.1 Flash 1017 processed 3.04T tokens and rose 345% against the day before. The prior Flash snapshot still ranked second that day at 2.62T. GLM 5.3 Flash, MiMo-V2.6-Flash and Hy4 preview filled the next Chinese slots before Claude Opus 5.5 and GPT-6 Luna.
That day board lines up with DeepSeek's Oct 17 model update, which we covered from DeepSeek's own model card and API docs. The OpenRouter chart is the usage follow-through: developers routed a large share of new traffic to the 1017 snapshot once it appeared.
Why this matters, and what it does not prove
OpenRouter is a routing layer popular with developers and coding agents. Its rankings page states that the charts show traffic through OpenRouter, count tokens and requests rather than users or spend, and do not score accuracy. OpenRouter notes that "a higher token total shows how much a model is used, not which model is best for a task." The Benchmarks section on the same page, which uses the Artificial Analysis Intelligence Index, was led by Claude Opus 5.5 when we checked on Oct 19.
The practical read for builders is narrow but clear. On this marketplace, Chinese models from DeepSeek, Z.ai, StepFun, Xiaomi and Tencent are where a large share of routed work currently lands. US labs still hold substantial request share through Google and OpenAI, and Anthropic's requests rose 25% week over week on the Oct 12 chart. The Chinese lead is strongest in the token ranking, where the entire top five is Chinese.
Businesskorea's report ties the shift to price and agents: because agents call models many times per task, small price gaps add up. It also reports that DeepSeek had five outages in October alone. OpenRouter's rankings page does not report outages. Two related stories sit outside these rankings: the US DHS order to remove DeepSeek from federal civilian agency IT, and China's proposed export controls on frontier AI models.
Sources
- OpenRouter LLM Rankings (usage data through Oct 18, 2026; market share for week beginning Oct 12; weekly and daily token leaderboards)
- Chinese AI firms challenge U.S. rivals on price (The Dong-A Ilbo, Oct 19, 2026)
- Global AI Price War Intensifies as Chinese Models Gain Ground (Businesskorea, Oct 19, 2026)
