HomeArticle

Doubao Work, WorkBuddy and Qianwen Office are competing on the same track: models are hidden in the background, while access entrances are moving to the forefront.

Tech商业2026-11-18 11:38
Leading domestic office AI agents are collectively shifting to the multi-model aggregation track.

In mid-November, a subtle yet watershed turning point took place in China's office intelligent agent track.

On November 12, Doubao Work under ByteDance announced that it has integrated a number of third-party models including Kimi K3.1 Pro, GLM 5.3, DeepSeek V4.1 Flash, etc., and marked the "quota consumption multiplier" of each model on the model selection page, allowing users to weigh between task requirements and usage costs on their own. Four days later on November 16, Qwen Work under Alibaba launched 10 models at one go. In addition to the self-developed Qwen4-Max, third-party models including Kimi, GLM, DeepSeek, MiniMax and Step are all included. Enterprise administrators can also configure model access permissions by department, user group or member.

So far, Tencent WorkBuddy, ByteDance Doubao Work, Baidu Kuku AI, Alibaba Qwen Work — the four leading domestic office Agent products have all embarked on the same route of multi-model aggregation. More than two weeks ago, when Doubao Work took the lead in opening third-party models to enterprise customers, Qwen Work was still the "last adherent"; more than two weeks later, the adherent changed its course and pushed the opening intensity to the maximum at one time.

This is not a simple function update. When the four platforms all invite their competitors' models into their own products by coincidence, the competition logic of office Agent is shifting from "whose model is stronger" to a proposition closer to the retail and platform economy: whoever controls the task entrance will hold the pricing power and user relationship.

Behind "Model Freedom": Not generosity, but an inevitable choice

To understand this collective shift, we need to first clarify a basic fact: the tasks faced by office Agent are inherently heterogeneous.

When an ordinary employee asks the Agent to complete an industry research report, the process may cover five links: data retrieval, data cleaning, financial analysis, trend deduction and PPT generation. Data retrieval requires long context throughput and fast response, financial analysis requires multi-step reasoning, and PPT generation returns to content organization and format processing. If the most expensive flagship model is called throughout the whole process, the computing power consumption will get out of control quickly; if we blindly pursue low cost, the reasoning error in the intermediate link will be amplified along the chain, and the rework cost will be higher instead. The "single-point lead" of model capability is easily diluted in front of the task chain.

At the same time, the change rhythm on the model supply side is also accelerating. One model may temporarily take the lead in long text processing, another has more solid Chinese understanding, and the third achieves extreme cost performance in reasoning cost. Capability advantages rise and fall like tides, and no one can stay ashore continuously in all dimensions. For the office Agent platform, betting on a single self-developed model is equivalent to binding the product experience to a variable that can never be constant.

The deeper pressure comes from the competition pattern. Tencent WorkBuddy opened the selection of third-party models such as DeepSeek, GLM and Kimi to users as early as its launch in March 2026, and then successively accessed the flagship models of Stepfun and Xiaomi. Data from Analysys shows that in September 2026, the monthly visits of WorkBuddy reached 85.34 million, while that of Doubao Work was 30.39 million, and the former was about 2.8 times of the latter. When the first mover has established user habits with "model freedom", if the latecomer continues to adhere to the exclusive supply of self-developed models, what it loses is not only one model call, but all the possibilities for users to entrust work tasks to this platform.

Therefore, although the turn of Qwen Work was tortuous — it briefly accessed GLM and DeepSeek in August, quietly went offline three days later, and returned with greater intensity three months later — it was almost inevitable. Alibaba finally chose the latter between the application demand of self-developed models and the competitiveness of platform products. This is not a denial of Qwen's capability, but a reconfirmation of the essence of office Agent competition: what users want is not "which company's model to use", but "get the work done at a reasonable cost".

From "Selecting Model" to "Routing Model"

Multi-model aggregation solves the problem of "having choices", but what really tests the platform's engineering depth and business wisdom is "how to choose".

Doubao Work has put forward a notable design at this step: the transparency of quota consumption multiplier. The consumption standards of all models are unified under the same quota system. The lower the multiplier, the less platform quota is consumed for the same task, and users can intuitively compare the cost difference during model selection. Before that, a common problem in the enterprise-level AI market was the opaque billing rules: some models priced input and output tokens separately, some charged by the number of calls, and the conversion between different units was so complicated that ordinary users could not judge the actual cost before the task started, and often found that the budget was exceeded when the bill came out. The multiplier marking pulls the cost decision-making power out of the black box and returns it to the person using the task.

But this is only the first step. Manually selecting models by users is still an action that requires judgment and experience. A more fundamental trend is emerging: model routing.

The so-called model routing essentially automatically matches the "good enough and cheap enough" model for tasks, instead of calling the most expensive one by default. Enterprise AI budget is undergoing a structural transformation: the extensive call of "default to the strongest" in the past is being replaced by "task-based hierarchical scheduling". Cases disclosed by Palantir show that by switching some tasks from strong models to lightweight models, the reasoning cost can be reduced by up to 97%; the AI token usage of construction company McCarthy Building decreased by about 60% year-on-year, mainly from model scheduling optimization. These figures point to the same direction: when model capabilities are gradually leveled, scheduling capability is becoming the core variable that determines the economy of AI applications.

Tencent WorkBuddy has built in "Auto mode", which automatically selects the appropriate model according to the task type without manual judgment by users. Microsoft's update to Copilot at Ignite 2026 also follows the same logic: Auto routing dynamically selects models and reasoning investment in the background, balancing quality, speed and cost, and employees only need to tell Copilot what they want to complete. A survey of BT150 enterprise executives by Constellation Research gives a refined summary: "Use the cheapest model and architecture that can reliably meet business needs".

This means that the competition of office Agent is entering a deeper level. The number of accessed models and the length of the menu are no longer the core barriers — as long as the business negotiation is in place, the model menu can be replicated within a few weeks. The real barriers lie in: whether the platform has accumulated enough real task data to train and calibrate the routing strategy; whether it can identify the dividing line between "good enough" and "must be strong" in every link of the task chain; whether it can convert cost savings into user-perceivable pricing advantages. Model aggregation is the entrance, and routing capability is the moat.

Who pays for "Model Freedom": Entrance locking and revenue reconstruction

If we pull the perspective from the product layer to the business layer, the financial logic of this multi-model movement will be clearer.

The way Doubao Work opens third-party models to enterprise customers is worth examining in detail: when customers use DeepSeek, Kimi or GLM, the AI points in the Doubao Work account will be deducted directly, and there is no need to sign another contract with the model manufacturer. When enterprises choose other companies' models, the money is still paid to ByteDance's product. QwenWork's enterprise permission management further structures this logic: administrators assign model access permissions by department and role, and the platform becomes the unified settlement layer and governance layer for enterprise AI consumption.

The essence of this model is to lock the exclusivity of the Agent entrance with the openness of model selection. Office Agent charges seat fees, and an enterprise customer often has dozens to hundreds of accounts, with ARPU an order of magnitude higher than C-end AI subscriptions. More importantly, what is precipitated in the office scenario is the user's task records, workflows and context — once a team builds their daily workflow on a certain Agent, the migration cost is far more than simply "changing a model". The platform gives up the exclusivity of model calls in exchange for the occupancy of the working environment.

For model manufacturers, the implication of this pattern is dual. On the one hand, office Agent has become an important distribution channel for model capabilities to reach enterprise users, and Kimi, GLM and DeepSeek have obtained exposure that is difficult to achieve through API sales alone by settling in the four major platforms. On the other hand, when the platform controls the task entrance and routing right, the position of model manufacturers in the bargaining structure will shift slightly backward. The platform tends to allocate call volume according to "task adaptation" rather than "brand loyalty", and the competition between models will further converge to cost performance and differentiated capabilities. As observed in the BT150 executive survey, enterprises are routing sensitive or predictable tasks to internal or smaller models, and the premium space for cutting-edge models is continuously compressed.

This also explains why model manufacturers themselves are migrating to subscription systems and enterprise-level intelligent agent solutions. The gross profit margin of Zhipu's API business just turned positive to 24.6% in the first half of 2026, while the reasoning cost decreased by about 80% in the past year. The profit margin of pure API supply has reached the top, and simply pursuing call volume can no longer support the valuation narrative. When the office Agent platform turns models into replaceable "intelligent supply", model manufacturers must look for thicker profits in the upstream application layer and delivery layer.

The value of the entrance is greater than the aura of the model

Looking back at the choices made by Doubao Work and QwenWork in the same November, they jointly confirmed one thing: in the office scenario, models are receding into the background, and Agent are moving to the foreground.

For users, this is certainly a kind of liberation. They no longer need to pay for the marketing narrative of "which model is the strongest", nor do they need to migrate the whole set of working environment due to the change of model preference. The task is handed over to the Agent, the model selection is handed over to the routing, and the cost calculation is handed over to the transparent quota system. Office AI has changed from a product that requires "faith" to a tool that can "calculate accounts".

But for the industry, this means that the shift of competition focus is faster than most people expected. The arms race of model capabilities will not stop, but it is becoming a background competition; what is contested in the foreground is the task entrance, workflow embedding depth and cost governance capability. WorkBuddy's first-mover advantage does not lie in how strong its Hunyuan model is, but in that it was the first to let users get used to "freely switching models in one workbench". What Doubao and QwenWork are catching up with is not the number of models, but the product container that makes the number of models meaningful.

When the four platforms put the same batch of models on the same menu, the menu itself is no longer a source of differentiation. The real question is: whose kitchen can turn these ingredients into a table of dishes that is the most stable, the most affordable and the most suitable for Chinese office scenarios. Model freedom is just the beginning.

This article is from the WeChat official account "Tech Business", written by Hai Yi and Ding Yi, and published with authorization from 36Kr.