DeepSeek Leads a Clean Sweep: Chinese Models Take All Five Top Spots in a Global Weekly Token Ranking

DeepSeek Leads a Clean Sweep: Chinese Models Take All Five Top Spots in a Global Weekly Token Ranking

Four Chinese labs, one Western absence in the top five

For the week of October 5–11, Chinese large models processed 86.74 trillion tokens on the developer routing platform OpenRouter, up 50.96% week over week, according to calculations by National Business Daily. US models handled 18.97 trillion tokens in the same week, up 17.1%, against a global platform total of 172 trillion tokens, itself up 3.61%. Every one of the five most-used models that week was Chinese: DeepSeek V4.1 Flash took first place with 41 trillion tokens, up 60%; Zhipu's GLM 5.3 Flash came second at 12.2 trillion, up 26%; Xiaomi's MiMo-V2.6-Flash third at 11.5 trillion, up 19%; StepFun's Step 5 Preview fourth at 8.84 trillion; and Tencent's Hy4 preview fifth at 8.28 trillion, up 27%. It was the 24th consecutive week that Chinese models led US models on this measure.

The platform's own chart agrees, with DeepSeek nearly three times clear

OpenRouter's published weekly rankings, with data through October 13, show the same shape: DeepSeek V4.1 Flash at roughly 40.5 trillion tokens, Step 5 Preview at 15.1 trillion, and GLM 5.3 Flash at 12.9 trillion. A separate week, September 7–13, put Chinese models at about 61.2 trillion tokens against 21.8 trillion for US models — close to three to one — according to figures from Yicai relayed by TrendForce, part of a 20-week run.

Read the caveats before the headline: whose numbers are these?

The figures measure OpenRouter traffic only, not the AI market as a whole, and each outlet aggregates differently — National Business Daily sums the models that appear on the leaderboard, while OpenRouter's own ranking counts everything routed through its API. Peter Walker, OpenRouter's head of insights, told CNBC that Chinese open models released this year can now handle advanced agentic use cases, coding especially, in a way that was not true in late 2025. The platform also stresses that its newer compliance features fail closed: if no eligible in-region provider can serve a request, OpenRouter returns an error rather than routing it elsewhere.

The discount is the mechanism, and cached tokens show it best

Third-party pricing compiled in October puts DeepSeek V4.1 Flash at $0.15 per million input tokens and $0.60 per million output tokens off-peak. OpenAI's GPT-6.1 Sol is listed at $2 and $10; Anthropic's Claude Opus 5.5 at $4 and $20. Run a million input tokens plus a million output tokens and DeepSeek costs $0.75, against $12 for GPT-6.1 Sol and $24 for Opus 5.5. Cached input reads on DeepSeek V4.1 Flash fall as low as $0.003 per million tokens off-peak. Chinese models are variously described as 10 to 50 times cheaper per token than US offerings, and up to 90% cheaper in some cases.

Agent workloads, not chatbots, are inflating the totals

Agentic traffic — long context, tool calls, retries on coding tasks — accounts for more than 60% of platform volume and is the most price-sensitive and token-hungry segment there is. Venture investor Tomasz Tunguz found 84% of tokens processed on OpenRouter run through non-frontier models. Chinese models have taken over 30% of US-originating token consumption on the platform every week since February 8, 2026, peaking at 46%; in early 2025 that share was about 4.5%. CNBC-reported usage data showed Chinese models at 57% to 67% of all OpenRouter tokens in the week of September 14, up from 6% to 13% in February, with 55% on Vercel in August against 11% in January, and 67% across the 82 countries OpenRouter groups as the Global South.

Volume is not revenue, and that gap is the real story

An Epoch AI estimate found DeepSeek captured only about 8% of customer spending on its models accessed through the platform. Open-weight models can be served by third-party inference providers, so much of the billing lands with cloud and inference vendors rather than the labs that trained the models. Jefferies cites OpenRouter figures showing Chinese models processed 76.12 trillion tokens in the week ending September 27, against 4.37 trillion in late April, while the top four US providers handled 14.2 trillion that September week. Volume up roughly seventeenfold, revenue nowhere near matching it.

Quality is converging even as the volume gap narrows

Bloomberg Intelligence senior analyst Robert Lea wrote that after DeepSeek shipped V4.1 Flash in September, top Chinese models trailed US rivals by only about 3% on benchmark scores, down from roughly 9% in May and 15% earlier in the year — and argued that continued improvement points to further share gains. The ratio of Chinese to US weekly usage, calculated by National Business Daily, hit 14.1 times in late July and has since fallen to about 3.5 times. On October 15, Intel CEO Lip-Bu Tan described DeepSeek, Kimi and GLM as "good enough," pointing to low token costs and fast adoption.

What to watch: open weights land, and stealth models vanish

StepFun's Step 5 Preview is a sparse MoE with 600 billion total parameters and about 27 billion active per token, a 1 million-token context window and vision input. It scores 44 on the Artificial Analysis Intelligence Index, which the company says places it in the global open-source top three, and its weights were due to be released on October 15 at $1.00 per million input tokens, $2.70 output and $0.05 for cache reads. Meanwhile the anonymous model that briefly topped the chart — Space Bunny Alpha — dropped off the leaderboard entirely in the week covered here. Whether the top five stays an all-Chinese clean sweep is the signal to track next.