TechWeb
【TechWeb】依据第三方AI大模型聚合平台OpenRouter放出的最新测算数据,10月19日至10月25日的统计周期内,全球AI大模型总调用量达到188万亿Token,总体环比前一周增长14.63%,全球AI情形的商业化落地热度还在持续走高。其中,中国厂商推出的AI大模型周总调用量达93.54万亿Token,环比增长18.14%;同期美国AI大模型周总调用量为22.47万亿Token,尽管环比增速达到29.44%,但总体调用规模仅约为中国的四分之一。这已是中国大模型周调用量连续第二十六周超过美国,稳居全球首位,国内AI大模型的to B、to C全情形落地普及度,形成了对海外市面的显著规模长处。
上周全球调用量排行榜前十(按Token计)
榜单总排名前五的席位全部被中国AI大模型包揽,没有任何一款海外头部机型挤进前五。具体来看,DeepSeek V4.1-Flash-1017以33.6万亿Token登顶第一,环比暴涨801%;DeepSeek V4.1 Flash(旧版本)排在第二,周调用量18万亿Token,环比下滑45%;MiniMax M3.1 Flash (free)冲上第三,周调用量13.7万亿Token;智谱GLM 5.3 Flash位列第四,周调用量13.2万亿Token,环比增长8%;腾讯Hy4 preview连续第五周留在前五,周调用量8.97万亿Token,环比增长19%。海外机型中排名最高的是第六位的Claude Opus 5.5(8.04万亿Token),小米MiMo-V2.6-Flash则跌至第七。
值得注意的是,上榜的海外模型并非缺席,Anthropic的Claude Opus 5.5环比增长35%,OpenAI的GPT-6 Luna也浮现在前十,但与中国模型在量级上的差距已拉开一个身位。
V4.1-Flash-1017:一次版本更新引爆调用增量
榜首的V4.1-Flash-1017表现最为突出。该版本由DeepSeek于10月17日官宣更新,以9月10日发布的V4.1 Flash为基础,借助基于自反馈的强化学习训练工艺,重点改善推理、编程、自动化及视觉任务表现——在DeepSeek自家公布的基准中,Terminal-Bench 4.0得分从31.2提高至42.3,ProgramBench从20.3提高至29.1,Humanity’s Last Exam从36.8提高至42.6,并新增对Search与Fetch接口的兼容(5元/千次)。售价维持不变,每百万输入Token 0.15美元、输出0.60美元,模型权重以MIT协议开源。
同属DeepSeek的旧版V4.1 Flash周调用量环比下滑45%,两者此消彼长,恰好展现了典型的"版本迁移效应"——使用者正快速涌向表现更优、售价不变的新版本,33.6万亿Token的增量中相当部分来自自家旧版本的使用者迁移,叠加官方网页端、App与API的同步切换,进一步推高了OpenRouter上该模型的路由流量。本次更新在DeepSeek所称的多个测试框架下达到了"表现-费用帕累托最优",以更高性价比触发需要爆发,说明市面对实用型、可负担的模型能力存在巨大缺口。这也印证了近期榜单的普遍规律:登顶者往往靠的是售价策略与版本迭代节奏,而非单纯的模型能力代差。
冷静看待:调用量第一不等于工艺全面领先
OpenRouter的数据常被用作观察AI大模型格局的关键窗口,但其统计口径的局限需要正视。OpenRouter本质是一家API聚合网关,其榜单只统计借助该网关中转的Token量——大量模型借助厂商官方API、Azure、Bedrock、谷歌Vertex或私有化部署给出服务,这部分调用不会被完整记录。有行业测算指出,OpenRouter每周的Token调用量仅占全球总量的很小比例;平台使用者结构上,美国使用者占比近半,英语内容占绝对主导。因而,"连续26周超越美国"反映的是平台渠道内中国模型的受欢迎程度与售价竞争力,不宜直接等同于中美AI产业的综合实力比较——OpenAI、Anthropic等美国头部模型的实际调用量有相当部分走自有渠道,未进入统计。
从表现维度看,据斯坦福大学《2026年AI指数报告》,截至2026年3月,中美顶尖模型之间的差距仅为2.7%,双方已多次交替登顶表现榜单。中国模型的核心长处在于"足够好的表现+显著更低的成本":有美国研究机构测算,Anthropic头部模型平均每项任务成本约为DeepSeek V4.1 Flash的28倍,叠加开源策略,让全球开发者可以低成本调用甚至自主部署。在东南亚,中国大模型借助本地化语言适配应对英文模型"水土不服"的痛点;在欧美,DoorDash、爱彼迎、西门子等企业为降低工艺成本纷纷接入中国模型;在非洲,越来越多开发者把中国开源模型当作本地创新的"工具箱"。
但另一维度,调用量领先无法直接等同于产业收入或利润领先——榜单第三的MiniMax M3.1 Flash以免费模式冲入前三,本身就说明计费策略会显著影响这一指标。Token量也并非衡量价值的统一尺度,简单文本生成与复杂编程、多步Agent推理的成本和收益不可同日而语。流量规模上的长处能否转化为可持续的商业闭环,仍是中国AI产业需要回答的难题。
从这个意义上说,188万亿Token的周调用纪录、93.54万亿的中国份额、801%的版本爆发增长,共同勾勒出的是中国大模型在应用生态活跃度和成本竞争力上的真实领先;而在榜单之外,关于高端算力供给、前沿表现突破与商业化盈利路径的竞赛,仍在继续。
本文来源:https://tech.sina.cn/2026-10-26/detail-iniwpwit5437910.d.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
TechWeb
[TechWeb] According to the latest estimates released by the third-party AI large model aggregation platform OpenRouter, during the statistical period from October 19 to October 25, the total number of global AI large model calls reached 188 trillion Tokens, a total increase of 14.63% week-on-week, and the commercialization adoption momentum of global AI continues to rise. Among them, the total weekly call volume of AI large models launched by Chinese vendors reached 93.54 trillion Tokens, up 18.14% week-on-week; during the same period, the total weekly call volume of U.S. AI large models was 22.47 trillion Tokens, and although the week-on-week growth rate reached 29.44%, the overall call scale was only about one quarter of China's. This marks the 26th consecutive week that China's large model weekly call volume has surpassed that of the United States, firmly occupying the top spot globally. The adoption penetration of domestic AI large models across both to-B and to-C scenarios has formed a significant scale advantage over overseas markets.
Top 10 Global Call Volume Rankings Last Week (by Token)
The top five spots in the overall rankings were entirely claimed by Chinese AI large models, with no overseas flagship models breaking into the top five. Specifically, DeepSeek V4.1-Flash-1017 topped the list with 33.6 trillion Tokens, a staggering 801% increase week-on-week; DeepSeek V4.1 Flash (old version) ranked second with a weekly call volume of 18 trillion Tokens, down 45% week-on-week; MiniMax M3.1 Flash (free) surged to third place with a weekly call volume of 13.7 trillion Tokens; Zhipu GLM 5.3 Flash ranked fourth with a weekly call volume of 13.2 trillion Tokens, up 8% week-on-week; Tencent Hy4 preview remained in the top five for the fifth consecutive week, with a weekly call volume of 8.97 trillion Tokens, up 19% week-on-week. The highest-ranked overseas model was Claude Opus 5.5 in sixth place (8.04 trillion Tokens), while Xiaomi MiMo-V2.6-Flash fell to seventh.
It is worth noting that overseas models were not absent from the list—Anthropic's Claude Opus 5.5 grew 35% week-on-week, and OpenAI's GPT-6 Luna also appeared in the top ten, but the gap in scale with Chinese models has already widened considerably.
V4.1-Flash-1017: A Version Update Triggers a Surge in Call Volume
The top-ranked V4.1-Flash-1017 performed most outstandingly. This version was officially announced by DeepSeek on October 17 as an update, based on the V4.1 Flash released on September 10. Leveraging a self-feedback reinforcement learning training process, it focuses on improving performance in reasoning, programming, automation, and vision tasks—in benchmarks published by DeepSeek itself, the Terminal-Bench 4.0 score improved from 31.2 to 42.3, ProgramBench improved from 20.3 to 29.1, Humanity's Last Exam improved from 36.8 to 42.6, and it added compatibility with the Search and Fetch APIs (5 yuan per 1,000 calls). The price remains unchanged: $0.15 per million input Tokens and $0.60 per million output Tokens, with model weights open-sourced under the MIT license.
The old version DeepSeek V4.1 Flash saw its weekly call volume drop 45% week-on-week. The rise and fall of the two versions perfectly illustrates the typical "version migration effect"—users are rapidly flocking to the new version with better performance and unchanged pricing. A significant portion of the 33.6 trillion Token increase comes from the migration of users from the company's own old version. Combined with the synchronized switch across the official website, App, and API, this further boosted the routing traffic for this model on OpenRouter. This update achieved what DeepSeek describes as a "performance-cost Pareto optimum" under multiple test frameworks, triggering an urgent surge in demand with higher cost-effectiveness, indicating that the market has a huge gap for practical, affordable model capabilities. This also confirms a general pattern in recent rankings: the top-ranked models often rely on pricing strategies and version iteration rhythms, rather than pure generational gaps in model capability.
A Sober Perspective: Top Call Volume Does Not Equal Comprehensive Technology Leadership
OpenRouter's data is often used as a key window for observing the AI large model landscape, but the limitations of its statistical methodology need to be acknowledged. OpenRouter is essentially an API aggregation gateway, and its rankings only count the Token volume routed through this gateway—many models provide services through vendors' official APIs, Azure, Bedrock, Google Vertex, or private deployments, and these calls are not fully recorded. Industry estimates suggest that OpenRouter's weekly Token call volume accounts for only a small proportion of the global total; in terms of platform user composition, U.S. users account for nearly half, and English content is absolutely dominant. Therefore, "surpassing the United States for 26 consecutive weeks" reflects the popularity and pricing competitiveness of Chinese models within the platform's channels, and should not be directly equated with a comparison of the comprehensive strength of the U.S. and Chinese AI industries—a considerable portion of the actual call volume of top U.S. models such as OpenAI and Anthropic goes through their own channels and is not captured in the statistics.
From the performance dimension, according to Stanford University's "2026 AI Index Report," as of March 2026, the gap between the top Chinese and U.S. models was only 2.7%, and the two sides have repeatedly alternated at the top of performance rankings. The core advantage of Chinese models lies in "sufficiently good performance + significantly lower cost": a U.S. research institution estimates that the average cost per task of Anthropic's top models is about 28 times that of DeepSeek V4.1 Flash. Combined with the open-source strategy, this allows global developers to call at low cost or even deploy independently. In Southeast Asia, Chinese large models address the pain point of "acclimatization issues" of English models through localized language adaptation; in Europe and the United States, companies such as DoorDash, Airbnb, and Siemens have integrated Chinese models to reduce technology costs; in Africa, more and more developers are using Chinese open-source models as a "toolbox" for local innovation.
However, from another dimension, leading in call volume cannot be directly equated with leading in industrial revenue or profit—the fact that MiniMax M3.1 Flash, ranked third on the list, broke into the top three with a free model itself shows that pricing strategies significantly affect this metric. Token volume is also not a unified measure of value; the costs and benefits of simple text generation versus complex programming and multi-step Agent reasoning are not comparable. Whether the advantage in traffic scale can be converted into a sustainable commercial closed loop remains a difficult question for China's AI industry to answer.
In this sense, the weekly call record of 188 trillion Tokens, China's share of 93.54 trillion, and the 801% version growth spurt jointly outline China's large models' genuine leadership in application ecosystem activity and cost competitiveness; while beyond the rankings, the race regarding high-end computing power supply, frontier performance breakthroughs, and commercialization profitability paths continues.