Do you really believe WorkBuddy has 20M MAU?

Summary

• WorkBuddy does not have 20M MAU. Analysys’s reported 20.97M is the “number of interactions between users and AI-native office-agent platforms on PC,” not MAU or UV, and not even PV in the usual sense; 庄明浩 estimates the real user count could be just one-tenth—or as little as one out of several dozen—of that figure.
• The more credible public statement comes from 姚顺雨: WorkBuddy now has “more than 1M daily active users.” The claimed 12M-13M DAU, 庄明浩 says, came from a typo in an investment-bank record, according to 汪天凡; WorkBuddy’s mobile app did not officially launch until July 18 and cannot explain data through the end of June.
• Chatbots are giving way to Agents. OpenAI recast Codex as the new ChatGPT and added “Work,” while the old chat product was relegated to “ChatGPT Classic”; one week after the rename, Codex weekly active users jumped from 6M to 10M. 庄明浩 calls it a product-level “bait-and-switch,” with the signal that “Code was first, now everything else.”
• China’s desktop-Agent market remains in an early-adopter phase. Products began appearing in late March 2026, with launches concentrated in April-May (months 4-5); most have been in operation for no more than 1 quarter, while usage remains concentrated on PC and access is still not frictionless. Using 100M-200M general white-collar computer users as the ceiling and “10%-20% at most” penetration today, 庄明浩 puts the market at 10M-40M people.
• QuestMobile’s data puts another constraint on the estimate. China has 499M AI-native app users and 644M users of in-app AI plugins, but only 18M users of PC-client AI applications—and that figure includes Chatbots such as Doubao, Qianwen and Yuanbao. In other words, actual desktop-Agent users are likely below the 18M count, rather than any single product already having 20M MAU.
• The final benchmark is not users but money. The market is asking less often how many users ChatGPT or Claude Code have, and has stopped taking token consumption on faith; attention has shifted to ARR. Yet the monetization paths available to Chinese AI applications—productivity subscriptions, consumer payments and To B services—are precisely the models that have repeatedly struggled over the past 10-plus years. “Will AI be different this time?” remains unanswered.
• 庄明浩’s conclusion is not to write off the product but to correct inflated metrics. WorkBuddy’s current public description is “more than 1M daily active users,” but the debate over the number should not obscure the product shift. “Let’s not worry about the rest for now; just start using all kinds of tools.”

Deep dive

1. Models iterate monthly, but the real regime change is happening at the product entry point

• From GPT-3.5 in late 2022, to o1 in 2024, DeepSeek and Claude in 2025, and Grok 5, Opus 5, GPT-5.6 and Kimi K3 in 2026, 庄明浩’s sense is that the pace is now “clearly impossible to hold back.” The industry has started tracking progress by the month; one running joke is that the US-China relay has compressed the cycle to “every half-day.”
• More important than yet another model release is OpenAI recasting Codex as the new ChatGPT and adding “Work,” while renaming the original chat entry point ChatGPT Classic. 庄明浩 describes the product reshuffle as “a bait-and-switch”（狸猫换太子）: the leading model company is using naming and default placement to declare that work-oriented Agents are the next phase.
• The data also shows the power of rebuilding the entry point: roughly one week after the rename, Codex weekly active users rose from 6M to 10M. It was the first time the show had cited “10M active users,” but it also raised a key question: given access costs and the barriers to use, how many of those 10M weekly users are actually in China?
• 庄明浩’s dividing line between eras runs from Chatbots that can only converse and Copilots that assist humans to systems that use tools, call different models and run multiple Agents in parallel. If he had to identify one moment, he would mark the release of Claude Code in 2025 as the “Claude Code moment.”

2. Code first, then work: Harness and evaluation loops become new infrastructure

• The Claude client added a parallel Cowork button alongside Chat; Codex now offers both Code and Work. 庄明浩 summarizes the trajectory in one line: “Code was first, now everything else.” Coding Agents validated the paradigm, and the same mechanics are now spreading across broader knowledge work.
• The next layer to emerge is Harness: tools, Skills, MCP, system prompts, models and runtime infrastructure together form the Agent middleware. The ecosystem not only determines how products are built but is beginning to steer model training in return; open-source attention in the first half of the year has therefore centered on Coding Harness, Personal AI Assistant and the conversion of Chatbots into Agents.
• AI product development is also moving from “first devise the perfect solution, then hand it to engineering” to an evaluation-driven climb: measure how well a product solves one task or problem, then iterate continuously to raise its score. 庄明浩 says this consensus formed “very, very quickly” in the Work Agent category.
• He uses Mario to explain the full stack: the model is like Little Mario; a mushroom grants Code, Web, Memory, Computer Use, Store and Skill Store; another flower supplies context scattered across different places, turning him into Mario with the ability to fire bullets.

3. Big tech’s collective bet does not mean desktop Agents have broken out

• China has seen an unusually crowded field: Tencent is advancing WorkBuddy as a “super project,” with 马化腾 reportedly almost never missing a product meeting; Alibaba has consolidated several Agents into Qianwen Office; ByteDance- and Doubao-related moves arrived in the same period, while 360 launched Nano Work. 庄明浩’s read is that the market has reached consensus on the Work Agent direction.
• Product heat and user penetration are not the same thing. “Lobster” and its variants only caught fire around the Spring Festival; most standalone Buddy, Work and Claw products appeared after late March, with launches concentrated in April-May (months 4-5). Most have been in operation for no more than 1 quarter.
• The use case itself limits diffusion: these products are primarily PC-based, and many are not even websites or apps. Even as vendors work to lower the barrier, they remain unintuitive for ordinary users. As of the end of July 2026, 庄明浩 thinks they “probably have not broken out”; penetration of 10%-20% is already the absolute ceiling.
• He funnels urban employment, the broad white-collar population and the historical scale of DingTalk, WeCom, Feishu, WPS and Tencent Meeting into an upper bound of 100M-200M people who can use desktop Agents. Applying 10%-20% early penetration gives a rough current user base of 10M-40M, or “the low tens of millions,” not 10M users for every leading product.

4. 20.97M counts interactions; three data points puncture the “20M MAU” claim

• Analysys’s ranking lists WorkBuddy at 20.97M “visits,” while media reports broadly rewrote that as 20M MAU. The catch is in the report’s footnote: “visits” means the number of interactions between users and the PC platform during the measurement period. A high-frequency user can interact several times a day, so the monthly figure should be divided by per-user frequency.
• 庄明浩 therefore warned media and analysts against using the report casually: “This is not MAU, not UV, and not even PV.” His rough conversion is that real UV or MAU could be one-tenth—or even one out of several dozen—of the ranking figure; total June interactions across all products tracked by Analysys amounted to only slightly more than 60M.
• QuestMobile imposes an even tighter constraint: AI-native apps account for 499M users, AI functions embedded in WeChat, Alipay and other apps for 644M, device-maker AI for 755M, and the PC web for 172M. But PC-client AI has only 18M users, a figure that also includes the client users of Doubao, Qianwen and Yuanbao.
• On the SimilarWeb measure, WorkBuddy and CodeBuddy together recorded 7.31M web visits in June. 庄明浩 says that number can broadly be treated as PV, but it still needs to be divided by frequency to arrive at UV. Separately, 汪天凡 said the purported 12M-13M DAU came from a typo in a record; 姚顺雨 told LatePost that WorkBuddy had “more than 1M daily active users.” With the mobile app only officially launching on July 18, it cannot lift the June figure—and “more than 1M” itself gives no precise range.

5. Users and tokens take a back seat; AI applications ultimately face the revenue test

• While working through the user numbers, 庄明浩 realized that the market has not seriously tracked ChatGPT’s user base for some time, and almost nobody is asking how many users Claude Code has. Without the Codex rename, the market might not have cared about its active-user figure either; attention has shifted from user growth to ARR.
• Token consumption is no more reliable an answer. “Nobody is chasing so-called token maxing anymore”; 1B versus 2B tokens does not directly establish commercial value. 庄明浩’s bearish conclusion is blunt: “Money is always the most naked standard.”
• The AI business models visible today are still direct monetization: productivity tools, individual subscriptions, and API- or seat-based To B services. Those are precisely the areas where Chinese internet companies have struggled to build durable businesses over the past 10-plus years. After Doubao had been charging users for a little over 1 month, 庄明浩 judged the result “not ideal,” even though it may be the domestic AI application best positioned to test paid adoption.
• Token Plan appears supply-constrained and continues to raise prices, but “Is this volume big enough? I don’t know.” The episode leaves the final question open: whether “AI will be different” remains unanswered. Rather than let inflated MAU figures dictate the narrative, the better move is to start using the Agent tools.