Hands-on comparison of three AI office apps: Doubao Work, Qwen Office, and WorkBuddy
Reporters from The Beijing News Shell Finance conducted a series of hands-on task tests on three AI office apps—Doubao Work, Qwen Office, and WorkBuddy—using the Doubao 2.1 Lite0921, Qwen3.87-Max, and Hy4 preview models, respectively. In the long-horizon task from the "2026 AI Office Sector Research Report," Doubao Work produced 10,000 words in 20 minutes 46 seconds, WorkBuddy produced 25,000 words in 21 minutes 18 seconds, and Qwen Office took over 1 hour to produce 50,000 words but consumed about 1,800 tokens, roughly 60 times that of the other two, requiring the reporter to top up 78 yuan to purchase a membership. In the multimodal tests, Doubao Work's poster and video had the highest degree of completion, and the 3D model generated by WorkBuddy was the closest to the original figurine, but when generating a comic it mistook its own icon for a teal pig face.
Why read this
AI score 65
A hands-on comparison of three AI office apps on the same set of tasks, detailing specific differences in long-horizon tasks, multimodal generation, and token consumption.
Topics
Body · AI translation
Since August of this year, a trend has become increasingly obvious—all kinds of "lobster apps" are bidding farewell to the "Warring States era" of rival warlords dividing the territory. The lobster apps under Tencent, Alibaba, and ByteDance that were originally launched by different internal teams as "horse-racing" projects are gradually being consolidated into three single brands: WorkBuddy, Qwen Office, and Doubao Work.
Among them, Doubao Work produced a blockbuster-style instant-feel advertising video from just a single photo of earphones, but the 3D model generation task "heavily altered" the real object photo; Qwen Office produced the richest report content on long-horizon tasks, but also "burned money" to the point that the reporter had to top up to save the situation; WorkBuddy had the highest degree of openness, but also had moments of "misjudging," mistaking its own icon for a "teal-green piglet face" when generating a comic.
The big tech companies collectively "withdrew their troops," but the three giants' products still have ecosystem barriers
Within two or three months, the big tech companies almost simultaneously completed the consolidation of their organizations and products.
Alibaba was the first to make its move. On August 3, Qwen Office loudly launched its public beta, and news emerged from inside Alibaba that the organizational consolidation of three teams—Wukong, MuleRun, and QoderWork—had been completed, establishing the Qwen Office business unit. Shortly after the Shell Finance reporter registered for Qwen Office, they received multiple follow-up calls as well as SMS and DingTalk messages, showing how much Alibaba values the product's user experience.
On August 25, ByteDance released Doubao Work. Before that, ByteDance had also completed the merger of the Doubao, Feishu, and Volcano Engine teams internally. The teams behind products such as ArkClaw, TRAE, Coze, and Feishu OpenClaw were merged in whole or in part into Doubao Work. On September 15, ByteDance CEO Liang Rubo appeared at the 2026 Feishu Future Unlimited Conference and Doubao Work Kickoff Meeting, showing up for the first "combined" appearance of Feishu and Doubao Work.
By comparison, Tencent "sat firmly on the fishing platform." WorkBuddy is currently the most impressive-performing app in the "lobster war," and Pony Ma said in his speech on Tencent's 2026 Q2 financial report that it had achieved "breakthrough user growth." On September 24, Tencent officially shut down QClaw, which had originally competed internally with WorkBuddy in a "horse race," and it is reported that the relevant business and team of the QClaw product center have also been adjusted into the department to which WorkBuddy belongs.
This collective move sent a signal: domestic big tech companies have gradually passed through the exploratory period of multi-product horse racing and begun to contract resources, seeking a product for their own ecosystems that can serve as a stable backstop.
Judging from the interfaces of Doubao Work, Qwen Office, and WorkBuddy, there is not much difference among the three. Screenshot
During testing, the Shell Finance reporter discovered ecosystem barriers. For example, images in WeChat chat boxes can be smoothly dragged into WorkBuddy, but cannot be dragged directly into Doubao Work and Qwen Office. Over the past month, when the reporter communicated with product managers of the above AI applications, they often emphasized the interconnectivity of their own applications, such as DingTalk for Qwen Office, Feishu for Doubao Work, and Baidu Netdisk for Baidu Dazi. To a certain extent, this means that AI office applications are still an extension of the big tech companies' respective ecosystems.
Writing research reports: Doubao drew illustrations for its report, Qwen wrote 50,000 words
Starting in September this year, reporters from The Beijing News Shell Finance successively conducted a series of task tests on three AI office applications: WorkBuddy, Qwen Office, and Doubao Work. Among them, WorkBuddy used the Hy4 preview model when executing tasks, Qwen Office used the Qwen3.87-Max model, and Doubao Work used the new Doubao 2.1 Lite0921 model. These models are all the most advanced models under Tencent, Alibaba, and ByteDance, ensuring that the tasks were executed by the big tech companies' "original teams."
To test the long-horizon task capabilities of the agents, the reporter separately asked the above three applications to generate a "2026 AI Office Track Research Report." Among them, Doubao Work dispatched four "research sub-agents" and completed the report writing in 20 minutes and 46 seconds, totaling 10,000 words; WorkBuddy retrieved the reporter's previous work history from the workspace, "instantly understood" the task, and completed the report writing in 21 minutes and 18 seconds, totaling 25,000 words; Qwen Office was the most rigorous, confirming the delivery format, length and depth, and report focus with the reporter in the form of "single-choice + multiple-choice questions" before the task began, and the final task process took more than 1 hour, producing a 50,000-word draft.
From left to right are screenshots of the research reports generated by Doubao Work, WorkBuddy, and Qwen Office.
Judging from the results, all three AIs produced in-depth research reports that met the requirements, including detailed data analysis, comparison standards, and source lists. In terms of form, Doubao Work's report had the most beautiful appearance, producing not only tables but also bar charts and capability radar charts; WorkBuddy provided many unique observation angles, such as "users are not choosing the strongest AI, but choosing where the data they already have is accumulated"; while Qwen Office provided the most detailed and content-rich research report, covering the two-orders-of-magnitude gap between the monthly active user volume of AI office applications and their parent apps, as well as a detailed summary of the attitudes of executives at various big tech companies toward AI office applications.
However, this report, which had the largest word count, was not cheap. For this task, Doubao Work consumed 31.91 tokens, and WorkBuddy's consumption was also around 30 tokens, which could be covered by the free quota. In comparison, Qwen Office's Qwen-3.8Max consumed about 1,800 tokens in one go, nearly 60 times that of the former. Because the tokens were consumed too quickly, the reporter had to top up 78 yuan on the spot to buy a one-month membership.
WorkBuddy's comic generation mistook its own icon, while 3D generation staged a "comeback"
In the multimodal capability test, the generation capabilities involved comics, advertising posters, videos, and 3D models.
A reporter from Shell Finance sent screenshots of the desktop icons of Doubao Work, Qwen Office, and WorkBuddy to the above AI applications and entered the prompt "Design a comic in which three little people wearing the above icons are fighting each other."
From left to right are screenshots of the research reports generated by Doubao Work, Qwen Office, and WorkBuddy.
The results showed that Doubao Work generated a comic image with the icons "on top," while Qwen Office and WorkBuddy understood the reporter's intent more thoroughly, turning the icons directly into characters, making the comics more vivid. However, while the other two generated their icons correctly, WorkBuddy "mistook" its own logo. When the reporter reviewed its thinking process, it found that WorkBuddy described its own icon as a "white piglet face logo on a teal background," which may have caused the subsequent image generation error.
The thinking process of WorkBuddy generating the comic. Screenshot
Advertising posters generated based on a photo of wired earphones, from left to right, generated by WorkBuddy, Doubao Work, and Qwen Office.
The results showed that, whether for posters or videos, the deliverables of Doubao Work, which relies on the Seedance series models, had a more polished quality. In this test, there were obvious differences in the video effects generated by the three AI applications. WorkBuddy's live-action model moved naturally and also had storyboards, but the entire video had no sound; Qwen Office's video understood that "product" was the top priority, with the camera gradually pushing toward the earphones, but it lacked the sense of variation from storyboards; Doubao Work, by contrast, had both lighting changes and multi-layered storyboards of the model holding the earphones, plugging them in, and wearing them to enjoy music, and the voiceover did not feel out of place, giving it the highest degree of completion. The reporter noted that the three videos shared the same flaw: possibly due to a lack of material on wired earphones, the earphones generated in all three videos did not match the real product.
For the 3D model generation test, which tests capabilities even more, a reporter from Shell Finance uploaded four photos of an anime girl figurine from different angles and requested, "Create a 3D model of this figurine so that I can rotate it in all views directly by dragging with the mouse in a browser."
The upper left corner is the original figurine photo, the upper right corner is the 3D model generated by WorkBuddy, the lower left corner is generated by Doubao Work, and the lower right corner is generated by Qwen Office.
The 3D models of Doubao Work and Qwen Office directly compressed the figurine image into a rough art style, among which Qwen Office's character image was even more rudimentary than Doubao Work's. WorkBuddy, however, generated a 3D model almost identical to the original figurine. In terms of the control feel when selecting the camera, both Doubao Work and WorkBuddy were very natural, while Qwen Office showed a certain degree of lag.
During the actual test, the reporter found that the current capabilities of AI applications had improved significantly compared with when "Lobster" became wildly popular earlier this year. When faced with difficulties encountered during task execution, such as the computer environment lacking necessary components, "Lobster" previously often could not solve them independently and needed to ask the user for help; in this actual test, however, the above AI office applications showed various "creative solutions" during their thinking processes.
In a transcription task for a medical consultation recording, Doubao Work encountered a situation in which security software blocked the network, making the task difficult to complete. While looking for a solution, it discovered transcription material that WorkBuddy had completed earlier. Doubao Work decided to "generate a standardized document based on the existing transcription content and synchronize the relevant information."
The decision-making process of Doubao Work when network requests were blocked during task execution. Screenshot
However, as the field moves from "Lobster everywhere" to the three strong contenders consolidating, the AI office track is gradually moving past the "can it be used" stage, but hallucinations have not disappeared. Under homogeneous tasks, ecosystem migration costs, token consumption, and other factors will all become implicit criteria for users' choices, and the balance between capability and cost will determine who can retain heavy users.
According to data obtained by a reporter from the Beijing News Shell Finance from Xsignal, in August this year, the top three applications by monthly active users in the AI office track across APP and WEB were WorkBuddy, Baidu Dazi, and Coze, indicating that the integration effects of the aforementioned giants have not yet fully manifested in the short term. Whether the future will be "separate regimes," a "tripartite balance," or whether some product can "unify the world" is currently unknown.
A data overview of monthly active users on the APP and WEB ends of AI office applications in August 2026. Produced by Luo Yidan, reporter of Beijing News Shell Finance. Data source: Xsignal
Produced by Du Xiaotong and Ren Wanqing, reporters of Beijing News Shell Finance. Data source: Xsignal
Reporter contact email: luoyidan@xjbnews.com
1 International | Last recording before the medical plane lost contact exposed! It plunged 3,000 meters after taking off from Bermuda 31,000; 2 Trending in Beijing suburbs | Yongning Market was booming during National Day, with 56,500 visits to the intangible cultural heritage market 18,000; 3 Comrade Lu Shengdao passed away 18,000; 4 US media: Apple confirms that some US versions of iPhone 18 Pro Max have malfunctions and need replacement devices 17,000; 5 A suspected murder case involving US forces stationed in Japan triggered protests from Japan; the US side promised full cooperation with the investigation 9,945
The US aircraft carrier "Eisenhower" will return to the United States to undergo an accident investigation 16 minutes ago; Money in the home of a telecom fraud leader in northern Myanmar was so plentiful it grew moldy! Real case disclosed 17 minutes ago; In response to safety hazards such as "landing visas" for motorcycle commercial photography, Chongqing launched a special rectification campaign 20 minutes ago; Single-day passenger flow at Hengqin Port exceeded 150,000, setting a new historical high 33 minutes ago; Railway in the Yangtze River Delta is expected to send more than 3.8 million passengers in recent days, with return passenger flow increasing 33 minutes ago
Central Meteorological Observatory: Cold air moves south, rain stops and skies clear in most parts of the country; winds in the eastern sea areas are relatively strong; Waitlisted train numbers show they can be submitted but pop-up windows repeatedly appear; 12306 responds: caused by too many submissions; From "living map" directions to "iron fist" rectification, Capital Airport police protect safety in all respects; Police forces follow the flow of people, and Haidian Public Security uses "one park, one policy" to weave a dense holiday safety net; Djokovic comes from behind to beat Zverev, achieving 32 consecutive wins at the China Open