文|Tech商业
2026年夏天的北京,办公场景正悄然生变。以前上班族开机第一件事是启动飞书、钉钉或企业微信,如今越来越多的人习惯先点开一个AI图标——用它来撰写周报、梳理会议纪要,甚至随手做一页PPT。推动这一转变的,是四家科技巨头在短短两个月内的高频布局:6月豆包启动付费,8月千问办公、百度库库AI以及字节豆包工作相继面世。
功能清单固然亮眼,可没人愿意把底下的账本摆到台面上。AI每跑一次调用,都是真金白银的消耗,然而过去两年间,国内用户早已被"全民免费"的模式培养出了一个根深蒂固的认知:AI就应该免费。2026年,正是这四款产品头一回共同直面这道商业难题的年份。
先把各自的底牌看清楚。接下来这张表并非财务报表,而是呈现了四款产品的出发点、切入点,以及它们各自计划如何实现盈利——
表格中描绘出四条风格迥异的商业化路线:有的依赖订阅分级,有的依靠交易分成,有的凭借工具融入工作流程,有的则借助矩阵式覆盖。不同路线的背后,折射出各家面对"AI是否应该收费"这一棘手问题时,各自的解题思路与信心所在。
豆包:规模既是铠甲,亦是命门
豆包堪称国内AI产品中把规模做到极致的代表。据QuestMobile统计,其月活跃用户数从3月的3.45亿攀升至6月的3.82亿,同比增幅达172.1%,位列中国移动互联网总榜第16位;6月25日——也就是专业版订阅推出的第二天——日活跃用户数飙升至1.78亿的历史高点,成为当时国内仅有的日活破亿的独立AI应用。花旗调研显示,豆包在国内AI应用领域的用户渗透率已高达79%。
然而规模既是铠甲,亦是命门。2026年3月,豆包日均Token调用量已达120万亿,6月更是冲破180万亿大关,相比2024年5月刚上线时暴涨逾1500倍。对于一款主打免费的产品而言,用户每增长一分,成本账单就加厚一层——规模非但没能摊薄支出,反而将成本体量持续放大。
6月24日,豆包正式上线专业版,设置连续包月三档:68元、200元、500元。外界争议颇为激烈,但用户的逻辑站得住脚:在这之前,豆包给大多数人的印象就是"免费且足够好用"。两年的全民免费早已把价值预期牢牢锚定,而豆包主打的卖点恰恰是"更强模型能力、更多调用次数"——这种能力既能横向比较,也完全可以忍住不买。
让字节真正看到破局希望的,是另一条产品线。视频生成模型Seedance系列月营收已突破10亿元,毛利率在70%左右。它售卖的并非底层能力,而是能够直接交付的成品。
豆包的解题思路在8月25日揭晓——正式推出面向生产力场景的Agent产品"豆包工作"。该产品配备独立的桌面客户端,同时在飞书侧边栏开放入口,支持撰写文档、制作表格、生成PPT,并可在授权范围内调取企业内部的飞书文档、会议纪要及知识库;定价采用个人与企业双轨制,个人版延续专业版的三档定价结构,企业版则按席位销售并支持加购AI用量包,发布时即赠送30天订阅权益。
为了给这款产品提供支撑,字节在短短一个月内完成了一轮罕见的高频组织架构调整:7月底飞书产品团队并入豆包、飞书GTM团队整合至火山引擎,8月24日TRAE与扣子团队也整体并入豆包体系。
豆包坐拥近4亿月活用户,但C端用户并不会自然而然地转化为企业客户。一旦踏入企业门槛,组织架构、权限管控、采购流程等关卡都在前方等着。豆包工作面临的真正考验并非技术能力,而是如何将用户的"聊天习惯"转化为对"工作流的深度依赖"。
千问:将AI嵌入一笔交易之中
当豆包还在为"如何让用户付费"而苦恼时,千问根本就没碰这道难题——它直接将AI变成了交易链条的前置步骤。
千问的增长速度在中国AI发展史上实属罕见。QuestMobile数据显示,其移动端月活从3月的1.66亿增长至6月的1.67亿,同比飙升5792.9%,排名由去年12月的第五位跃升至第二位。阿里巴巴财报透露,截至2月千问全端(App+Web+PC)月活已突破3亿。
这场冷启动并非白手起家。2025年11月26日,千问与夸克AI浏览器实现深度整合,而夸克在PC端的装机量已达1.1亿,这批存量用户顺理成章地转化为千问的首批流量来源。
真正的分水岭出现在1月15日:千问App全面打通淘宝、支付宝、淘宝闪购、飞猪、高德等阿里全系生态。春节期间,它协助用户完成近2亿笔订单,约1.3亿用户通过千问的智能体首次体验了AI购物。
大多数AI助手的终点是一段文字建议,用户仍需自行操作落地;千问的终点则是一笔已经成交的订单。它的收入并非来自订阅费用,而是源于交易本身——完全不需要说服用户"为AI买单"。
8月3日,千问补上了办公端的拼图:"千问办公"(QwenWork)启动公测,该产品由QoderWork、MuleRun、悟空三款产品整合而成,同步接入新旗舰模型Qwen3.8,并初步打通钉钉IM;8月18日又接入了企业微信,至此实现了对钉钉、飞书、企业微信三大协同平台的全面覆盖。其主打的差异化卖点名为"组织级Skill"——资深员工在完成一项复杂任务后,可将完整流程一键封装为技能,并向整个团队共享。
千问是四款产品中唯一不需要回答"用户凭什么为AI付费"这个问题的选手,因为它将这道难题,不动声色地转移到了已有的交易链条上。
元宝与WorkBuddy:腾讯的双线布局
这组对照极具分析价值。2026年,腾讯先后推出了两款方向完全迥异的AI机型。
元宝的开局可谓声势浩大。春节期间腾讯投入10亿元红包补贴,试图复刻当年"微信红包"的破圈效应。短期表现确实亮眼:元宝日活突破5000万,月活触及1.14亿,除夕当日日活高达4504万,主会场累计抽奖参与量超过36亿次,用户通过"创作"栏执行的AI任务也突破了10亿次。
然而热度消退的速度超出预期。据QuestMobile数据,元宝独立App月活在3月便滑落至5735万,到6月继续下探至4984万,未能站稳亿级门槛。摩根士丹利的追踪更为直观:春节假期结束后,元宝日均使用时长相比1月缩水32%,三周之后日均使用均值仅余6.5分钟。
6.5分钟意味着什么?它暴露了元宝并未真正嵌入任何日常工作流程——用户只是"偶尔想起才问一句",而非"开工首件事就打开它"。红包换来的只是下载量和一次性的尝鲜体验,而非使用习惯。
腾讯面向C端还有第三款机型正处于灰度测试阶段:微信原生AI助手"小微",搭载自研的WeLM模型,已接入公众号AI总结、朋友圈AI点评、AI帮写、扫一扫等16个高频使用场景。不过从实际体验来看,它目前仍停留在"60分助手"的水平——群聊总结和信息提炼几乎不产生幻觉,简单短链路任务勉强够用;可一旦指令超过三步,比如要求它"点一份50元以内的外卖",便会陷入卡顿。
WorkBuddy则走出了一条截然不同的路径。它于2026年3月9日开启公测,截至6月PC端访问量达到2097万次(易观分析数据),位列国内同品类之首,且超过了排名第二的字节TRAE IDE国内版(1279万次)与第三名阿里QoderWork(788万次)两者的总和。花旗统计口径显示其跨平台月活突破2000万、日活超过1300万;此外,依据AI机型榜·桌面榜7月数据,WorkBuddy桌面端月活为1115.23万,排名AI办公智能体品类第一。
它为何能留住用户?关键在于三大支柱:其一,接管本地文件,支持直接上传、读取、整理乃至修改本地文档;其二,产出结构化成果,仅需一句话即可生成带目录的会议纪要、待办表格以及完整的复盘PPT,且交付物支持编辑;其三,权限边界清晰,本地访问与权限管控并行——这正是它能够切入政务、国企领域的前提条件。
技术底座也提供了坚实支撑。混元Hy3于7月正式发布,据腾讯透露,在WorkBuddy、CodeBuddy等实际应用场景中,Agent任务成功率从72%跃升至90%。这18个百分点的提升直接转化为用户体感——用户能够明显感知到"它更频繁地把任务真正做完了"。
颇为耐人寻味的是其商业化策略:企业版虽已于5月上调价格,但管理层却明确表态"未给团队设定商业化KPI,目前仍处于投入期"。在四款机型当中,它是唯一在增长阶段主动选择暂不变现的产品。
腾讯通过两款机型印证了同一个结论:AI产品的用户留存源于工作流嵌入,而非补贴驱动。元宝坐拥10亿红包和微信社交关系链,WorkBuddy则毫无流量加持,最终胜出的却是后者。
库库AI、百度搭子与秒哒:百度的三张底牌
文心是四款中处境最为微妙的一个。作为最早入局的基础大模型之一,它在这一轮模型竞赛中的排名确实出现了下滑。李彦宏在8月18日的财报电话会议上明确表态,要推动文心大模型"重新回到基础大模型的第一梯队",百度近期也引进了1997年出生的复旦大学博士孙天祥,由其担任基础模型研发部(BMU)负责人。
不过百度真正的着力点,是在机型层面构建产品矩阵,而非在模型层面死磕单一突破口。
2026年8月完成AI办公产品线整合之后,百度形成了三款定位分明的机型:百度搭子,定位专业办公智能体平台,聚焦复杂任务的连续处理,7月MAU环比大增1063.79%,位列增速榜首位,桌面端MAU达674.30万;库库AI(GenFlow),面向通用办公,主打"跨设备、跨时空"的全天候自动化能力,办公MAU超过2500万,百度宣称其位居AI办公赛道第一;秒哒,专注无代码应用生成,可通过自然语言直接生成可用应用,沙利文数据显示其市场份额为33.4%,排名行业第一。
这与腾讯的策略恰好相反:腾讯走的是单兵路线(WorkBuddy一款打天下),百度则采取矩阵覆盖(三款机型分别覆盖不同颗粒度的办公需求)。矩阵模式的优势在于覆盖面广、总用户基数大,但代价是品牌分散、用户认知成本偏高。
有一组数据最能折射百度的机型处境:2026年第二季度,百度核心AI新业务收入为125亿元,占到一般性业务收入的50%,但细分来看——智能云基础设施贡献73亿元(同比增长50%,其中GPU云增长283%),AI原生营销服务贡献26亿元,而真正机型侧的"AI应用"仅贡献25亿元,同比仅微增3%。
换言之,百度AI收入的重头来自算力售卖,而非机型本身。这与其余三家面临的本质上是同一道难题,只不过百度将其摊开呈现在了财报之中。
百度另一张王牌在政企市场:自研昆仑芯、文心大模型与飞桨框架构成了全栈闭环。百度透露,目前100%的平台关键性银行及80%的央企选择了百度智能云;据智能超参数统计,2026年上半年国内五家主要AI云厂商的大模型相关中标总额约20.44亿元,其中百度智能云以13.85亿元位居榜首,占比接近六成。
办公桌成为新战场
将四家的举措排列在同一时间轴上,发展脉络便一目了然:3月9日WorkBuddy公测,6月24日豆包专业版,8月3日千问办公,8月14日库库AI,8月18日千问办公接入企业微信,8月25日豆包工作。四款机型全部落子完毕,赛道格局基本确立。依据7月数据,国内AI办公桌面端总月活约为3000万,WorkBuddy(1115.23万)与百度搭子(674.30万)两者合计占比约六成,而参照传统办公软件的用户规模测算,AI办公领域仍存在约20倍的增长空间。
办公场景是少数能让 AI 独立实现闭环的领域。聊天、搜索和内容生成往往只停留在“给建议”的层面,而办公任务却有着明确的交付物——一份文档、一张表格或一份 PPT,只要交付物能够被验收,其价值就能被量化定价。这一逻辑在 AI 编程领域早已跑通:代码结果能够进行自动化验证,用户能立刻感知到效率的提高,办公场景只不过是这一能力的自然外溢。
更为核心的变化在于,办公入口正经历着一场迁移。以往员工上班的第一步是开启文档、IM 软件或邮箱,而未来他们可能会优先启动智能体,让 AI 去调用各种工具。一旦智能体成为了用户打开电脑后的第一站,像钉钉、飞书、企业微信这样的平台就会从前台入口退居幕后,转化为被 AI 调用的底层能力模块。入口的转移本质上代表着权力的重新分配,没有任何玩家愿意在这场变革中掉队。
上半年的竞争焦点在于谁的使用者更多,而下半年的较量则变成了谁能真正把活干完。混元 Hy3 将 WorkBuddy 的任务成功率从 72% 提升至 90%,这表明当前全新的核心衡量指标已变成“任务成功率”——使用者根本不在乎你的模型多大,他们只在乎这次的活到底干完没有。这一指标将引发行业的新一轮洗牌:模型竞赛并未终结,只是从台前转向了幕后,只有那些在长链路任务上表现更稳定的模型,其下一代机型才能掌握真正的定价权。
本文来源:https://www.jiemian.com/article/15066688.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
文|Tech商业
The summer of 2026 marks a subtle shift in Beijing's office landscape. No longer is the first step after booting up to launch Feishu, DingTalk, or WeChat; instead, an increasing number of workers are opening an AI icon — using it to draft weekly reports, organize meeting minutes, or even whip up a quick PowerPoint slide. This transition is driven by the rapid-fire deployment of four tech giants within just two months: Doubao launched its paid service in June, followed by Thousand Qwen Work, Baidu KuKu AI, and ByteDance's Doubao Work all in August.
While the feature lists are undoubtedly impressive, none of these players are eager to share the underlying cost sheet. Every time an AI model is invoked, it represents a real financial expenditure. However, over the past two years, domestic users have been conditioned by the 'free-for-all' model, establishing a deep-seated belief that AI should be provided at no cost. The year 2026 is the first time these four products are simultaneously facing this commercial challenge head-on.
Before diving into the strategies, it's crucial to understand where each product stands and what their revenue generation plans are:
The table outlines four distinct monetization paths: some rely on tiered subscriptions, others on transaction commissions, and still others aim to integrate tools into existing workflows or leverage a matrix-style coverage. These different approaches reflect how each company has interpreted the difficult question of whether AI should be charged for, and the confidence they have in their solutions.
Doubao: The Power of Scale, But Also Its Achilles' Heel
Doubao is the poster child for scaling an AI product in China. According to QuestMobile, its Monthly Active Users (MAU) skyrocketed from 45.59 million in March to 49.41 million in June, representing a staggering year-over-year growth of 172.1%. It now ranks as the 16th most used mobile internet service in China. On June 25th, the day after the launch of the professional subscription, its Daily Active Users (DAU) peaked at a historic 178 million, making it the only independent AI application to break the 100 million user mark. Citi research shows that Doubao's penetration rate in the Chinese AI market has reached 79%.
However, this massive scale is a double-edged sword. In March 2026, Doubao was processing 120 billion tokens per day. By June, that number had exploded to 180 billion, a 1500% increase compared to its launch in May 2024. For a product that started life as a free service, user growth translates directly into exponential cost increases. The scale is not diluting the spending; it is amplifying it.
On June 24th, Doubao introduced its professional plan with monthly subscription tiers of 68, 200, and 500 RMB. The backlash was immediate. The general consensus among users was that the service was "good enough for free." Two years of free access have firmly anchored the value expectation at zero. The selling point ("more powerful models, more calls") is entirely optional — users can easily choose not to pay.
Where ByteDance sees a glimmer of hope is in its video product, Seedance. This specialized service is already generating over 1 billion RMB per month with a gross margin of roughly 70%. Unlike Doubao's chatbot, Seedance sells a tangible, usable product (videos), which makes the value proposition much clearer.
The strategy for Doubao became clear on August 25th with the launch of "Doubao Work." This new product is explicitly targeted at productivity and works as an agent. It features a dedicated desktop client and a side-panel entry within Feishu, allowing it to handle document creation, spreadsheet management, and PPT generation while accessing company-specific data like meeting minutes. It is priced using the same subscription model as the professional plan for individuals, with a new "seat" pricing structure for enterprises.
To support this pivot, ByteDance underwent a significant organizational overhaul. In July, the Feishu product team was merged into the Doubao team, and the Go-To-Market (GTM) team was merged into Volcano Engine. In August, the TRAE and Kouzi teams were also absorbed into the Doubao structure.
The challenge here is not just technical. While Doubao has nearly 400 million registered users, converting a "chat" habit into a "work pipeline" dependency is a different animal. The company has to figure out how to get users to pay for a feature that, for most of its life, was free.
Thousand Qwen: Embedding AI into the Transaction
While Doubao is grappling with how to get users to open their wallets, Thousand Qwen sidestepped that question entirely. It didn't wait for users to be ready to pay for AI; it inserted itself directly into the transaction process so that paying becomes a natural consequence of the interaction.
Thousand Qwen's growth has been phenomenal. QuestMobile reports its MAU grew from 16.6 million in March to 16.7 million in June, with a jaw-dropping year-over-year increase of 5792.9%. Alibaba's earnings report reveals that the total MAU (across app, web, and PC) surpassed 300 million by February.
This wasn't a cold start. In November 2025, Thousand Qwen integrated deeply with the Quark AI browser, which had an existing install base of 110 million on PCs. Those users became the initial seed for the chatbot.
The real game-changer came on January 15th when Thousand Qwen opened up to the entire Alibaba ecosystem: Taobao, Alipay, Taobao Flash Shopping, Fliggy, and GaoDe maps. Over the Chinese New Year, it assisted in nearly 200 million orders, introducing roughly 130 million users to AI-assisted shopping for the first time.
Most AI assistants end the interaction with a "here's your analysis," leaving the user to take the next step. Thousand Qwen doesn't stop there — it completes the order. Its revenue comes from the commission on the sale, not from a subscription fee.
On August 3rd, Thousand Qwen expanded into the office space with the beta release of "Qwen Work." This product aggregates several of Alibaba's internal tools (QoderWork, MuleRun, Wukong) and integrates directly with DingTalk IM. It also supports Enterprise WeChat. The unique selling point here is "organizational-level skills," allowing experts to package complex task sequences for the entire team to use.
The beauty of this model is that Thousand Qwen does not need to convince anyone to pay for AI. It simply makes it a part of the shopping cart experience.
Tencent: The Two-Pronged Approach
Tencent's strategy in 2026 offers a fascinating case study in contrast, launching two very different products: Yuanbao and WorkBuddy.
Yuanbao was launched with a massive marketing blitz during the Chinese New Year, backed by a 1 billion RMB red envelope subsidy. The results were immediate: DAU spiked to 50 million, and MAU reached 114 million. On New Year's Eve alone, the app saw 45.04 million users.
However, the post-holiday fallout was severe. By March, the MAU had dropped to 57.35 million, and by June, it was only 49.84 million. Morgan Stanley tracked the usage data and found that the average daily use time dropped by 32% after the holidays, settling into a meager 6.5 minutes per day.
This data is telling. A daily average of 6.5 minutes suggests that Yuanbao failed to integrate into any kind of daily workflow. Users downloaded it for the novelty of the free money, but once the promotion ended, they forgot it existed. It lacks the "stickiness" required for a sustainable business model.
Tencent's other horse in the race, WorkBuddy, tells a different story. Launched in March, it quickly dominated the PC AI agent market. By June, it recorded 20.97 million visits, far outstripping its competitors (ByteDance's TRAE IDE and Alibaba's QoderWork). Flower City reported a total MAU of over 20 million, and AI Model Ranking placed it firmly at the top of the productivity category for desktop users.
The secret to WorkBuddy's success lies in its practical functionality: * It can read and write local files, allowing users to edit documents directly within the chat. * It produces actionable results, such as full meeting minutes or complete slide decks, rather than just advice. * It has clear permission boundaries, making it acceptable for corporate environments.
Backing up WorkBuddy is the Hunyuan 3 model, which reportedly increased the task completion rate of the agent from 72% to 90%. This isn't just a marketing claim; it means that the robot actually gets the job done more often than not.
Interestingly, despite raising prices in May, Tencent has explicitly stated that there are no commercialization KPIs for WorkBuddy at the moment. It is still in an investment phase, suggesting the company is willing to wait for the technology to prove itself before forcing a revenue target.
The lesson from Tencent is clear: subsidies do not build habits; utility does. Users only stick around if the tool actually helps them do their job better.
Baidu: The Matrix Approach
Baidu's position is slightly precarious. While it was an early player with Wenxin, its ranking among foundation models has slipped. In the August earnings call, Li Yanhong admitted that they need to "return to the first tier" and promptly hired a new director of foundational model development.
However, rather than focusing solely on the base model, Baidu has chosen to build a "matrix" of specialized products. 1. **Baidu Buddy** (Dazuo): This is the premium product for complex, multi-step tasks. Its MAU jumped by 1063.79% in July, indicating strong growth potential. 2. **KuKu AI** (GenFlow): This is the general-purpose office assistant. With an MAU of over 25 million, it is Baidu's main driver for market coverage. 3. **Miaoda**: A specialized tool for turning plain text instructions into complete software applications (No-Code).
This fragmentation is a double-edged sword. On the positive side, it allows Baidu to cover every niche of the office market with dedicated solutions. On the negative side, relying on multiple products dilutes the brand recognition that a single, dominant tool (like ChatGPT) enjoys.
Looking at the financials, Baidu's AI revenue in Q2 was split roughly 50/50 between infrastructure (selling the chips and the cloud) and the actual applications. Crucially, the income from the AI "apps" (the chatbots) was only 25 billion RMB, which represents a microscopic year-over-year growth rate. This suggests that while Baidu has the technology, monetizing it at scale remains a challenge.
Baidu's strongest suit is not the consumer market; it is the enterprise and government sector. By owning the entire stack (the chip, the cloud, and the model), Baidu is able to provide a "secure" environment that is more appealing to state-owned enterprises. They currently hold 60% of the market share for AI contracts in this specific sector.
The Future: The Battle for the Launch Icon
The landscape for the next six to twelve months is shifting from a race for "who has the most users" to a race for "who has the most completed tasks."
The key metric moving forward will be the "task success rate." If an AI can complete a complex sequence of actions (read a document, extract data, update a report) without error, it becomes indispensable. If not, it is just a glorified search engine. This is where the battle for model supremacy will continue to rage, even if it happens in the background.
办公桌成为新战场
When the four companies’ moves are plotted on a timeline, the developmental trajectory becomes clear: WorkBuddy beta on March 9, Doubao Professional Edition on June 24, Qianwen Office on August 3, KuKu AI on August 14, Qianwen Office connecting to Enterprise WeChat on August 18, and Doubao Work on August 25. All four models have been deployed, and the market landscape is essentially established. Based on July data, the total monthly active users of domestic AI office desktop applications is approximately 30 million, with WorkBuddy (11.1523 million) and Baidu Companion (6.7430 million) accounting for about 60% together. Comparing with the user scale of traditional office software, there is still about a 20-fold growth potential in the AI office field.
The office scene is one of the few where AI can independently achieve a closed loop. Chat, search, and content generation often only offer suggestions, while office tasks have clear deliverables — a document, spreadsheet, or PPT. If the output can be accepted and verified, its value can be quantified and priced. This logic is already proven in AI programming: code results can be automatically verified, allowing users to instantly perceive efficiency gains. The office scene is merely a natural extension of this capability.
More critically, the point of entry for office tools is shifting. Employees once started their day with documents, IM software, or email; now they may prioritize launching an agent that manages everything for them. If agents become the first stop after turning on the computer, platforms like DingTalk, Feishu, and Enterprise WeChat will retreat to the background and become underlying modules called by AI. This shift represents a redistribution of power, and no player wants to be left behind.
The first half’s competition focused on user numbers; the second half is about who can finish the job. Hy3 improved WorkBuddy’s task success rate from 72% to 90%, indicating that the new core metric is “task success rate.” Users don’t care about model size — they just want the task done. This metric will reshuffle the industry: model races aren’t over, they’ve just moved backstage. Only models showing greater stability in long-chain tasks will have pricing power for their next iterations.
大模型 协同办公
微信公众号
微博
Source: https://www.jiemian.com/article/15066688.html. For copyright inquiries, email 13951431913#139.com or submit a comment on the right.
