8 Million Daily Users in Three Months vs a $2.5 Billion Run Rate: Doubao Work, Claude Code and Codex in the US-China Agent Race

Text
iHOLYSHITVISUAL / TECHIndustry changesiHOLYSHIT / ILLUSTRATION
iHolyShit illustration

On November 18, ByteDance's Doubao Work reported 8 million weekday daily users and over RMB 2 billion in enterprise ARR, three months after launch, while Tencent's WorkBuddy passed 10 million daily users the same day. Across the Pacific, Claude Code's run-rate revenue passed $2.5 billion months ago and Codex counts over 5 million weekly users. Both sides sell AI that does work — with opposite products, opposite economics and opposite model strategies.

On November 18, ByteDance's Doubao announced that Doubao Work, its productivity-focused AI product launched on August 25, had passed 8 million weekday daily active users in just three months, with annual recurring revenue from enterprise customers above RMB 2 billion. The same day, Tencent told its earnings call that its own office agent, WorkBuddy, had passed 10 million daily users and already counts 60% of China's top-100 companies as customers. China's office agents are now racing on scale. Across the Pacific, Anthropic's Claude Code and OpenAI's Codex dominate a different contest: not daily users, but how much each user pays.

Start with the product gap. Claude Code and Codex are coding agents that live in the terminal and the development environment, serving professional developers: they take requirements, read entire codebases, run tests and commit changes as part of an engineer's workflow. Doubao Work is a general office agent: users describe a task in natural language and hand over materials, and it breaks the task down, calls tools, and delivers research, content, data processing and even built applications. On November 16 it launched a Workbench feature covering 21 professional scenarios — finance, healthcare, education, creative design and more — with skills and tools pre-configured for each. One side serves people who write code; the other serves everyone who sits in an office. They are different races that happen to share a name.

The money gap is just as stark, and measured in opposite units. Anthropic disclosed in February that Claude Code's run-rate revenue had passed $2.5 billion, with enterprise customers generating more than half; third-party estimates put its annualized revenue per weekly active developer at roughly $1,250. OpenAI said in June that Codex had passed 5 million weekly users, with annualized revenue above $1 billion. Doubao Work's RMB 2 billion ARR is roughly $280 million, spread across 8 million daily users — most of them free, with the revenue concentrated in enterprise contracts. Gartner sizes the global enterprise AI coding-agent market at about $10 billion annualized. In short: the American products are collecting money, on professional depth and willingness to pay; the Chinese products are collecting users, on distribution and scenario breadth.

The deepest gap is the model strategy. On November 12, Doubao Work integrated third-party models — Kimi, GLM, DeepSeek and others — and labels each model with a quota-consumption multiplier, so users pick a model per task based on difficulty and cost. This is a move Anthropic and OpenAI structurally cannot copy: their products are vertically bound to their own models, and the model is the product. China's open-model ecosystem lets the platform layer exist on its own — route to whichever model is cheap and good, and earn money on orchestration and scenarios. The American moat runs the other way: frontier-model capability and developer mindshare. JetBrains surveys put Claude Code's workplace adoption among developers at 18% in January and 39% by mid-year.

In US-China terms, the two markets are decoupling. As this site has reported, Claude is fenced off from Chinese users behind layers of restrictions, while Chinese platforms are turning models into an open shelf at home, integrating each other's systems. In the short term the US leads on capability ceilings and paying markets; China leads on distribution, cost and enterprise rollout speed. Over time the two sides are converging on the same destination: in July, Codex was folded into the ChatGPT desktop app alongside a new Work mode, moving from coding toward general office work, while Doubao Work's Workbench and app-building push it from general office tasks toward professional depth. The real dividing line is not the products but the market borders: whoever first wins over users in third markets — Southeast Asia, the Middle East, Latin America — wins the second half.

iHolyshit review Doubao Work plugging in Kimi, GLM and DeepSeek is a play only the Chinese market makes possible: if your own model is not the strongest, turn every model into your inventory. That is a different faith from America's vertical integration. Eight million daily users against $2.5 billion in revenue is not the same weight class today — but Chinese internet companies have always played users first and money second. America is defending profit; China is seizing population. The tiebreaker is which side the third markets choose.

November 19, 2026

Share this article

What do you think of this story?

Anonymous feedback, up to 5 reactions per article.

Reader comments

Comment without signing in, or sign in to display your member name.

Discuss the technology and the story. Hate, politics and war, threats, abuse, personal accusations, sexual content, ads, links and personal information are not allowed.

0/500

No comments yet. Be the first to share.

Sources