Doubao has integrated models such as Kimi, GLM, and DeepSeek and adjusted quota multipliers to compete with WorkBuddy; competition in office agents is shifting from model integration to routing and architectural capabilities.Article author and source: AllWeatherTech
DouBao, under ByteDance, is beginning to allow users to access larger models from competing companies.
On November 12, DouBao Work announced an update to its model lineup, officially integrating third-party models including Kimi K3.1 Pro, GLM 5.3, and DeepSeek V4.1 Flash, in addition to its existing self-developed models, DouBao 2.1 Lite and Pro.
This update is also accompanied by a price adjustment.
Doubao Work will label the consumption multipliers for different models on the model selection page, allowing users to independently choose a model based on task complexity, response speed, and usage cost. Additionally, the platform is introducing a limited-time promotion that reduces the consumption multipliers for Doubao 2.1 Lite and DeepSeek V4.1 Flash, and offers new users a 30-day free subscription.
In fact, Doubao Work had already launched third-party model services for its team and enterprise subscribers on October 29. This latest update, which expands the model lineup and enhances promotional offers, further establishes the multi-model strategy as a key approach for Doubao Work to attract users.
The shift in DouBao's operations may stem from competitive market pressures.
Previously, Tencent WorkBuddy has integrated products from multiple model providers, including HunYuan, GLM, MiniMax, Kimi, and DeepSeek, and automatically selects the most suitable model through Auto mode based on routing policies.
WorkBuddy has indeed established a significant lead. According to the Q3 2026 Office Agent Market Report released by Analysys, in September this year, WorkBuddy achieved 85.34 million monthly visits, while DouBao Work recorded 30.39 million, making the former approximately 2.8 times the latter.
Amid WorkBuddy’s continuous expansion of its user base and its push to dominate the office agent entry point, Doudou Work is facing increasingly direct competitive pressure.
In fact, ByteDance is not inexperienced with multi-model products. Long before DouBao opened its platform to third-party models, TRAE—a ByteDance AI programming tool—had already adopted a multi-model strategy. TRAE Work offers users choices among models from various providers, including Seed, GLM, DeepSeek, Kimi, and MiniMax, and also supports users in connecting their own model services.
In terms of industry trends, Baidu has also adopted a similar open strategy. In August this year, Baidu officially renamed the general-purpose agent GenFlow, jointly launched by Wenku and Netdisk, to Kuku AI, and launched a standalone office client and enterprise version.
Among these, the model selection list in Kuku AI includes third-party models such as DeepSeek, GLM, Kimi, and MiniMax, in addition to Baidu’s proprietary Wenxin Large Model.
Qoder, under Alibaba, is no different. In April this year, the Qoder community version officially launched support for direct API Key integration, enabling connectivity with multiple model providers such as Alibaba Cloud Bailian, Zhipu, DeepSeek, Kimi, and MiniMax—yet Qwen Office has not yet opened up support for third-party models.
The difference between multi-model and single-model strategies lies in the distinction between Agent products and traditional AI chat tools.
For general chat products, the model’s knowledge base, reasoning ability, and generation quality often directly impact user experience. However, agents must handle more complex tasks, where a single task may involve information gathering, document analysis, code writing, data processing, and finally generating and verifying the outcome.
Different models perform differently at these stages and have varying costs.
For an office agent platform, while integrating the most powerful models is important, matching the right model to each specific task is becoming a new competitive factor.
More noteworthy than the integration of the Bido model is the advancement of model routing technology.
This update for Doubao Work primarily addresses the issue of users independently selecting models. Users can decide which model to assign tasks to based on each model’s characteristics and its multiplier for quota consumption.
However, manually selecting a model by the user is not the end of a multi-model architecture.
For example, an employee might ask the Agent to complete an industry research report, a process that could include gathering public information, organizing financial data, analyzing industry trends, and creating a presentation.
Among these, basic information organization may only require a model with faster response times and lower costs; for complex financial analysis or multi-step reasoning, a more capable model should be invoked; and for tasks such as format adjustments, the system can switch back to a lower-cost model.
Using the most expensive model throughout the entire process may lead to unnecessary computational waste; however, solely pursuing low cost could result in errors and rework due to insufficient model capability.
Model routing aims to strike a balance between task performance, execution speed, and usage cost.
This capability has already been integrated into the product architecture of office agents.
Tencent WorkBuddy's Auto mode supports automatic model allocation based on routing policies. In addition, Tencent Cloud has launched a standalone model routing service that assigns tasks to different models based on factors such as request intent, model load, and response latency.
All-day Tech noted that DouBao Work's Auto mode also adopts a small-amount free model.
By lowering the barrier to using Auto mode, Doubao Work has the opportunity to attract more users to try automatic model selection, gradually shifting them from manually choosing models to simply submitting tasks and waiting for results.
For the platform, executing more real tasks also helps continuously evaluate the capabilities, costs, and task completion effectiveness of different models, providing a basis for optimizing routing strategies.
From this perspective, Doubao Work may be seeking not only more users and task volume, but also the authority to manage models for future office agents.
As ByteDance, Tencent, and Baidu successively open their third-party models, the range of models available to each company’s office agents is becoming increasingly similar. Simply integrating more models is no longer sufficient to create a lasting competitive advantage.
This also means that competition among office agents is increasingly shifting toward the technological architectures built above the models.
One important capability is model routing—assigning the most suitable model to each task based on task complexity, execution performance, and token cost. However, from a more comprehensive Agent technology architecture perspective, model routing is just one component.
In recent years, Agent Harness has garnered industry attention by further encompassing task planning, tool invocation, context management, execution state maintenance, and error recovery—determining whether an agent can reliably transform model capabilities into consistent task outcomes.
Both jointly impact the commercial efficiency of office agents.
It is currently under close watch whether各家 will further attract users by engaging in price wars.