The AI Office War: ByteDance, Alibaba, and Baidu All Fired in One Week
Doubao Work, Qwen Office, and Dazi all moved in the same week of August 2026. The prices are brutal, and the model behind Alibaba's launch is 3x cheaper than DeepSeek. Here is what happened.
In the last week of August 2026, all three of China's big AI companies fired in the office productivity category. ByteDance launched Doubao Work on August 25, Alibaba consolidated around Qwen Office with a new model on August 26, and Baidu upgraded Dazi on August 27. Same week, same target: the AI agent that runs your office work. The prices are the story, and the model behind Alibaba's launch is the one to watch.
---
Doubao Work (August 25): ByteDance goes for the floor price
Doubao Work is ByteDance's office agent, built around Feishu (Lark). You give it a goal; it breaks the goal into tasks, operates your computer and browser, and can hand long jobs to a cloud computer you check from your phone. Its edge is context: with a Feishu account, it inherits your enterprise knowledge, chat history, documents, meeting minutes, and schedules, within permission limits.
The pricing is aggressive. Personal plans start at ¥68 per month (about $9.50), with Enhanced at ¥200 and Premium at ¥500. Enterprise pricing starts around ¥99 per seat per month billed annually. New users got a free 30-day subscription at launch. ByteDance also restructured around the launch, merging the Feishu product team into Doubao and folding the TRAE and Coze teams into the same organization, with Coze's 200+ connectors feeding Doubao Work's skill library. Our Doubao Work card has the full breakdown.
Qwen Office (August 26): Alibaba's answer, with a model that changes the cost math
Alibaba answered within 24 hours. Qwen Office, at ¥98 per month, is positioned directly against Doubao Work, and it runs on the new Qwen3.8-Flash, a model that is arguably the bigger news than the office product.
Qwen3.8-Flash is a sparse model: roughly 125 billion total parameters, but only about 6 billion active per token. The numbers attached to it are worth slowing down for. API pricing is about 1 RMB per million input tokens and 3 RMB per million output tokens ($0.15 and $0.47 internationally), roughly three times cheaper than DeepSeek-V4-Flash. Alibaba claims training cost dropped about 90% versus the previous generation. The office product itself has a "Standard Mode" that Alibaba says runs single tasks about twice as fast while consuming roughly 75% fewer tokens, covering about 95% of everyday office work.
The weights are open-source. That combination, open weights, 3x cheaper than the cheapest serious competitor, and a claimed training cost collapse, is what makes the model the more durable story: whatever happens to the office app, the price of capable inference just reset.
Baidu Dazi (August 27): the incumbent plays its hand
Baidu upgraded Dazi the next day, leaning on scale rather than price: 15 industry-specific suites and 96 skills, with 6.74 million monthly active users behind the platform. Dazi has been the quiet incumbent in the Chinese AI assistant space, and the upgrade reads as defense: we already have the users, here is the product depth.
What the war means outside China
For most readers of this site, none of these apps are usable day to day. They are Chinese-first, built around Feishu, DingTalk, and the domestic ecosystem. Two effects reach beyond China though.
First, the API price reset. When a 125B-parameter model prices at $0.15 per million input tokens with open weights, every Western provider feels the pressure on the next pricing round. The international numbers Alibaba publishes are not decorative.
Second, the product pattern. ByteDance and Alibaba are both selling the agent as the interface: goal in, done work out, apps operated underneath. That is the same direction Anthropic and OpenAI are pushing with computer use and connectors, and the Chinese products are shipping it at prices that make the comparison uncomfortable.
If you build on model APIs, watch Qwen3.8-Flash pricing and the open weights, because the cost floor just moved. If you evaluate office agents for an international team, study the Chinese products for architecture, not deployment. For the assistant comparison side of the market, our Perplexity vs ChatGPT vs Gemini comparison covers the Western landscape.
Q: Is Doubao Work cheaper than Qwen Office?
A: Yes. Doubao Work starts at ¥68 per month, Qwen Office at ¥98, and Tencent's WorkBuddy at ¥99. ByteDance is clearly pricing to win the entry tier, with a free 30-day subscription for early users during the launch period.
Q: What is Qwen3.8-Flash, and why does it matter?
A: It is Alibaba's new sparse open-weights model: about 125B parameters with roughly 6B active per token. API pricing is about 1 RMB per million input tokens ($0.15), roughly three times cheaper than DeepSeek-V4-Flash, and Alibaba claims about a 90% drop in training cost versus the prior generation. It is the strongest argument that capable inference keeps getting dramatically cheaper.
Q: Can I use Doubao Work or Qwen Office in English?
A: The products are Chinese-first: interfaces, ecosystems, and support target the Chinese market. The underlying models handle English content, but for an English-first office agent, Claude with Google Workspace connectors or ChatGPT are the practical choices.
Q: Which product is winning?
A: Too early to say. Doubao Work has the lowest price and deepest Feishu integration, Qwen Office has the cheaper underlying model and open weights, and Dazi has the largest existing user base. The week's real signal is direction: all three converged on the same agent-over-apps architecture in the same week.