DeepSeek Tops OpenRouter With 35.4 Trillion Weekly Tokens as Chinese Models Take Top Five

DeepSeek takes the first two places

DeepSeek's V4.1 Flash 1017 processed 35.4 trillion tokens in OpenRouter's weekly rankings, putting it ahead of its predecessor and giving Chinese developers all five leading positions. The platform's rankings, carrying usage data through October 26, 2026, place DeepSeek V4.1 Flash second with 17.4 trillion tokens.

Z.ai's GLM 5.3 Flash ranked third at 13.8 trillion tokens, followed by MiniMax M3.1 Flash's free version at 12.6 trillion and Tencent's Hy4 preview at 9 trillion. OpenAI's GPT-6 Luna also registered 9 trillion tokens at the displayed precision, but occupied sixth place.

The figures on OpenRouter's AI Model Rankings update the figures circulated for October 19 to 25, which put DeepSeek's newest model at 33.6 trillion tokens with growth of 801%; the official page lists 35.4 trillion and growth of 382%. The earlier figures also placed MiniMax third and GLM fourth, reversing their order on the official page.

DeepSeek's two leading variants together processed 52.8 trillion tokens in the official weekly table. Its newest version alone handled just over twice the volume of its predecessor, a striking shift following an update announced on October 17.

OpenRouter connects developers to models from multiple companies through one application programming interface, or API. Its rankings describe activity passing through that service. The broader AI market also includes companies' direct APIs, other cloud services, private installations and consumer applications.

What a token ranking measures

OpenRouter counts both prompt tokens, the material sent to a model, and completion tokens, the material it produces. A token is a small piece of text, sometimes a word and sometimes part of one. A long document or an extended conversation can contain many thousands of tokens.

A request is a different unit: it is a call asking a model to do something. One request might ask for a short translation, while another could include a large software project and ask for extensive changes. Both count as one request, but their token totals can be very different.

OpenRouter aggregates usage into daily buckets using Coordinated Universal Time. Its weekly table covers seven days ending with the latest complete bucket, while its monthly table covers a trailing period of 30 days. The date above the rankings identifies the newest usage bucket.

Model variants receive separate entries, including free versions, and requests marked private are excluded. OpenRouter measures traffic in tokens or requests, depending on the chart; users, spending and model capability require different measures.

A larger Chinese lead, with a narrower counting base

Figures circulated for October 19 to 25 put total OpenRouter activity at 188 trillion tokens, up 14.63% from the previous week. The same coverage put Chinese models at 93.54 trillion tokens and US models at 22.47 trillion, and described a run of 26 consecutive weeks in which Chinese models had processed more tokens.

The more detailed breakdown identifies the Chinese and US totals as usage among the top nine models. On that basis, Chinese models handled about 4.16 times the US volume, calculated from 93.54 trillion divided by 22.47 trillion. Chinese models occupied six of those nine positions.

The published growth comparison put Chinese volume up 18.14% and US volume up 29.44% against the preceding week. US models therefore grew faster in percentage terms during that period, while Chinese models retained the much larger token total within the listed group.

The grouping matters for understanding the competition. OpenRouter hosts models from more countries and providers than those two national categories, and its total traffic extends beyond the leading nine entries. The 188 trillion figure describes the platform's reported activity, rather than worldwide AI consumption across every service.

DeepSeek leads requests as well as tokens

DeepSeek also led OpenRouter's separate chart for shares of text requests in the week beginning October 19. It accounted for 26.5%, ahead of Google at 20.1%, OpenAI at 18.5% and Z.ai at 10.7%.

DeepSeek's lead over Google was 6.4 percentage points, and its lead over OpenAI was 8 percentage points. Google and OpenAI together accounted for 38.6% of requests, giving those two US developers a substantial presence even as Chinese models occupied the leading token positions.

Anthropic held 5.5% of text requests, Qwen 4.7%, MiniMax 2% and Tencent 1.8%. Mistral AI accounted for another 2%, while all other authors together represented 8.3%.

Request shares help explain a different part of the market from token totals. A provider serving many short exchanges can hold a large request share, while another handling fewer but much longer jobs can process more tokens. Developers choosing between models may care about both the number of interactions and the amount of work each interaction contains.

The releases behind October's reshuffle

DeepSeek announced V4.1 Flash 1017 on October 17, building on V4.1 Flash, which had been released on September 10. The update targeted reasoning, coding, automation and tasks involving images. The company also made the model's weights available, allowing others to obtain the numerical parameters used to run it.

MiniMax released M3.1 Flash on October 19, shortly before its free variant entered the leading weekly group. The sequence placed two fresh releases directly into a period of rapidly expanding platform traffic.

• September 10, 2026: DeepSeek released V4.1 Flash.
• October 17: DeepSeek announced the V4.1 Flash 1017 update.
• October 19: MiniMax released M3.1 Flash.
• October 19 to 25: The weekly period covered by the widely circulated usage figures.
• October 26: The newest usage date shown on OpenRouter's official rankings.

DeepSeek's earlier V4.1 Flash remained a major source of activity despite falling 41% in the official weekly table. The newer variant rose 382%. Their combined presence suggests developers were using both versions during the transition, although the rankings provide no breakdown of individual customers' switching decisions.

Daily and monthly tables tell different stories

DeepSeek V4.1 Flash 1017 also led the most recent complete day, processing 5.42 trillion tokens. GLM 5.3 Flash followed at 2.27 trillion, with the older DeepSeek V4.1 Flash close behind at 2.21 trillion.

The daily leader's volume was down 6% from the preceding day, even while it retained first place. Tencent's Hy4 preview rose 44% to 1.4 trillion tokens, and GLM 5.3 rose 56% to 892 billion, showing how quickly activity can move between models.

Across the trailing 30 days, the older DeepSeek V4.1 Flash remained first with 123 trillion tokens. GLM 5.3 Flash followed with 50.7 trillion. The newer DeepSeek model ranked fourth with 42.7 trillion, behind Space Bunny Alpha at 46.9 trillion.

Xiaomi's MiMo-V2.6-Flash illustrates the difference between those windows. It ranked ninth in the official weekly table at 5.32 trillion tokens, down 44%, but fifth over 30 days at 38.8 trillion, up 807% against the previous 30 days. A strong month can include a weaker final week.

Why coding and agents consume so much

MiniMax said M3.1 Flash improved performance in complex agent tasks, computer operation, software development and cybersecurity. It also said online inference speed had nearly tripled and usage costs had fallen by 70%. Inference means running a trained model to generate an answer or perform a task.

An AI agent works through a sequence of actions rather than producing only one reply. A coding agent might inspect files, propose an edit, run a tool, read the result and revise its work. Each step can send more information back to the model, increasing token consumption.

Earlier figures published on February 24 showed the same pattern around a previous generation of Chinese models. MiniMax M2.5 led that weekly table with 2.45 trillion tokens, followed by Kimi K2.5 at 1.21 trillion. Chinese models accounted for 5.3 trillion of the top ten's roughly 8.7 trillion tokens, a share of about 61%.

MiniMax M2.5's weekly usage had risen 197%, while GLM-5's rose 158% and Kimi K2.5's fell 20%. MiniMax said M2.5 exceeded 3.07 trillion tokens in its first seven days after release. Those figures use different periods, but together show that rapid adoption and sudden shifts between models preceded October's reshuffle.

Kimi K2.5 was designed to coordinate as many as 100 agents working in parallel. OpenRouter also identified growth in inputs containing between 100,000 and 1 million tokens around M2.5, connecting the earlier surge to the large amounts of context used in agent workflows.

Lower prices change the economics of repeated work

Prices published around the February releases show why developers running many repeated steps pay close attention to cost. MiniMax M2.5 and GLM-5 each charged $0.30 per million input tokens, compared with $5 for Claude Opus 4.6, about 16.7 times as much.

Output prices were $1.10 per million tokens for MiniMax M2.5, $2.55 for GLM-5 and $25 for Claude Opus 4.6. At those listed rates, generating 10 million output tokens would cost $11, $25.50 or $250 respectively, before any other charges or discounts.

October's releases continued that focus on operating costs. DeepSeek used different prices for busy and quieter periods, while MiniMax offered uniform pricing. Both approaches give developers ways to control the expense of applications that repeatedly send long conversations, documents or code to a model.

DeepSeek also released its DeepInfer inference framework, described as supporting both Nvidia hardware and Huawei Ascend processors. Its reported processing capacity was about 155 trillion tokens per cluster per day. That figure concerns the infrastructure running model services, a different operating measure from the traffic routed through OpenRouter.

Traffic and spending follow different patterns

OpenRouter's reported weekly token volume rose from 5.51 trillion in late October 2025 to 188 trillion in late October 2026. Dividing the two figures gives roughly 34 times the earlier level, illustrating how quickly the platform's workload expanded.

Task spending figures describe another side of that growth. General tasks accounted for 32.1%, agent workflows for 30.9% and coding for 26.7%. Together, those three categories represented 89.7% of the spending shown in that breakdown.

Claude Opus 5.5 led the classification spending ranking with a 10.3% share. Classification means assigning a label or category, such as identifying a customer's request, tagging a document or deciding which workflow should receive a message.

Developers' bills depend on prices as well as token counts. A cheaper model can handle a much larger volume while generating less revenue per token. Profit then depends on the cost of computing, operating the service and serving paying customers. The figures leave provider revenue and profit for this period unspecified.

Chinese AI deployment also extends beyond general language models. Meteorological departments in more than 40 countries have used the MAZU weather warning model through cloud services, with customized versions deployed in seven countries, including Pakistan, Ethiopia and the Solomon Islands. Such specialist applications show another way models reach users outside public API rankings.

Key Points

• DeepSeek V4.1 Flash 1017 led OpenRouter's official weekly table with 35.4 trillion tokens, followed by V4.1 Flash at 17.4 trillion.
• Chinese developers occupied the top five model positions, with GLM 5.3 Flash third, MiniMax M3.1 Flash's free variant fourth and Hy4 preview fifth.
• DeepSeek held 26.5% of text requests for the week beginning October 19, ahead of Google and OpenAI.
• The reported 26 week Chinese lead concerns usage on OpenRouter; the detailed national totals cited for October 19 to 25 cover the leading nine models.
• Coding, agent workflows and lower prices help explain the growth in repeated, lengthy model interactions.
• OpenRouter's rankings measure platform traffic, while spending, model capability and provider profits use different measures.