HomeArticle

Doubao has integrated third-party models, don't rush to praise this round of updates.

万点研究2026-11-17 18:03
On November 12, Doubao newly integrated a number of third-party models including Kimi K3.1 Pro, GLM 5.3, DeepSeek V4.1 Flash and more. Is this a sign that Doubao is becoming more open, or that its self-developed models are not sufficient enough?

On November 12, Doubao Work, the office intelligent agent product under ByteDance, announced the completion of product update. In addition to the two self-developed models, the original Doubao 2.1 Lite and Doubao 2.1 Pro, it has newly accessed a number of third-party models including Kimi K3.1 Pro, GLM 5.3, DeepSeek V4.1 Flash and others.

In September 2026, Doubao under ByteDance officially launched a new Agent product and brand "Doubao Work" for productivity scenarios, targeting the specific track of office Agent.

In this regard, two interpretations are the most prevalent in the industry: one is that "Doubao has become open", accessing third-party models and building a multi-model shelf; the other is that "ByteDance's self-developed models are not sufficient", and self-developed models have coverage boundaries.

Apart from the two statements, there is another issue that should not be ignored. Doubao has been designated by ByteDance as the main entrance of its AI strategy. For a product endowed with such a position, why does it access a number of third-party models in the specific battlefield of office Agent, and even directly show the quota consumption multiplier to users?

In a nutshell, this is neither a story of "open ecosystem" nor a story of "incompetent self-development". It means that after the entrance expands, someone must make the computing power account clear.

Accessing third-party models hides a computing power bill

To make this account clear, we must first clarify the expenditure of Doubao.

According to a Bloomberg report in September 2026, ByteDance completed the delivery of a USD 29.6 billion syndicated loan. The original planned financing amount was USD 20 billion, which was expanded to USD 29.6 billion due to enthusiastic subscription from banks. The interest rate is SOFR plus 68 basis points, which is the lowest level of overseas loan spread for Chinese technology enterprises in history, second only to the USD 40 billion bridge financing arranged by SoftBank for OpenAI.

Low financing cost indicates that the capital market still has a high credit pricing for ByteDance. However, no matter how low the interest rate is, it cannot change the fact that the borrowed money will eventually be repaid, and debt repayment ultimately depends on the profitability.

According to 36Kr reports, ByteDance's revenue in the first half of 2026 increased by 30% year-on-year to about USD 120 billion, with a net profit of about USD 20 billion. While its profitability remains excellent, the net profit margin has declined for three consecutive years, which was about 26% in 2023, dropped to 21% in 2024, and further fell to about 16.7% in the first half of 2026.

More notably, ByteDance is evaluating pushing its total capital expenditure for 2026 to a maximum of USD 70 billion, more than double that of 2025; even according to a more conservative caliber, AI capital expenditure has been revised upward to more than RMB 2 trillion. In short, the earned money is flowing to data centers, servers and computing power clusters at a faster speed.

Under this structure, the position of Doubao is very delicate. According to China National Radio, at ByteDance's first all-hands meeting in 2026, Liang Rubo, CEO of ByteDance, stated that the company set the keyword for 2026 as "climb the peak bravely", and made it clear that the "peak" in the short term is the Doubao/Dola assistant application.

For a product with high expectations, the usual script is to scale up first and then talk about monetization. However, ByteDance's capital expenditure rhythm and profit statement trend have compressed the time window for "talking about monetization later".

In June 2026, value-added services on the Doubao App were officially launched. Paid subscriptions start from 68 yuan, up to 500 yuan per month, and the quota of office tasks and expert mode of the standard package is more than 5 times that of the free version. The quota system of Doubao Work runs under the same subscription level as the paid framework of Doubao App, and the office scenario quota is deducted from the AI points that users have purchased.

In front of the hundred-billion-level capital expenditure, the more critical signal of Doubao's subscription revenue lies in commercialization. Doubao began to tell the market through the subscription and quota system that it is no longer just a cost center carrying traffic, but a product with a clear paid entrance and tiered pricing capability.

For Doubao, subscription is the first step. What really needs to be verified is whether users are willing to pay continuously for "consuming computing power according to tasks". The signal has been released, but financially, this gap cannot be diluted only by scale growth, and explicit management of inference cost at the product level is also required.

Accessing third-party models is not new in itself. From the first day, WorkBuddy has supported multiple models including Hunyuan, GLM, Kimi, DeepSeek, MiniMax and others, with Auto mode for routing; the model list of Baidu Kuku AI and Alibaba Qoder is also not closed. In the office Agent track, "multi-model" is no longer a differentiated advantage, and not accessing multi-models will constitute a competitive disadvantage.

Therefore, what is really worth analyzing about Doubao accessing third-party models is not "who it has accessed", but the access method: deducting quota according to the multiplier, no additional procurement required for enterprises, and direct deduction from the existing AI points.

Relevant industry insiders told Wandian Research that the real key of this access method is to turn "which model to use" from a technical issue into a cost and pricing issue. Doubao Work deducts quota by multiplier, no additional procurement is required for enterprises, and the deduction is directly made from the existing AI points, realizing explicit management of inference cost through product design.

The multiplier is not a decoration, Doubao puts computing power cost on the table

Explicit management, reflected on the product interface, is to let users see "how much the task will cost" when selecting models.

The lower the multiplier, the less the consumption, and the higher the multiplier, the higher the consumption; the preferential multipliers of Doubao 2.1 Lite and DeepSeek V4.1 Flash have been reduced to 1% of the standard multiplier for a limited time, with a clear direction: simple tasks are oriented to low-cost models, and complex tasks are left to high-multiplier models.

On the enterprise side, administrators can configure the available model scope by member, department or scenario, and the usage of calling third-party models is directly deducted from the AI points purchased by the enterprise, no need to purchase separately from Kimi, Zhipu and DeepSeek.

Why is this not redundant? Let's look at the price first. According to Shanghai Securities News, by the middle and late October, the output price per million tokens of Kimi K3.1 Pro standard version is 68 yuan. In contrast, according to Sci-Tech Innovation Board Daily, the output price of DeepSeek V4.1 Flash during idle time in the same period is 4 yuan per million tokens, and 8 yuan during peak hours.

For the same office Agent and the same task, switching from Kimi K3.1 Pro to DeepSeek V4.1 Flash, the unit price on the output side alone differs by 8.5 to 17 times. This price difference itself explains the function of multiplier labeling.

Marking the multiplier is equivalent to pushing the background price difference to the selection interface: tasks such as basic information sorting, format conversion, and simple summary can be completed with models priced at 4 to 8 yuan per million tokens output, there is no need to run models priced at 68 yuan per million tokens; complex analysis, long-link reasoning, and high instruction following tasks are worthy of using high-multiplier models.

The multiplier does not change the actual unit price of any model, but it turns the background price difference into a cost signal on the front end: what users see is not the abstract judgment of "which model is cheaper", but the immediate prompt of "how much quota this task will deduct". The unit price remains the same, but the decision-making basis has changed.

Another key fact: the single task cost of office Agent is not equal to "one question and one answer". According to foreign media reports, a large-scale empirical study released by Stanford Digital Economy Laboratory in April 2026 (based on benchmarks such as SWE-bench) found that the token consumption of agentic coding tasks can reach about 1000 times that of ordinary code Q&A, and the main cost comes from the input side: before each action of the Agent, it needs to re-read the task context and historical steps, rather than the output itself.

The laboratory's study also pointed out that there is a significant variance in token consumption between multiple runs of the same model on the same task. This means that the core uncertainty faced by users is not "which model is smarter", but "how much quota this run will deduct, which I cannot know in advance".

When the consumption of a single task fluctuates greatly due to the length of context, the number of retries and the tool calling path, "let's run it first" becomes a decision with financial consequences. This is exactly the function of Doubao Work marking the multiplier: it does not promise an accurate cost figure, but makes "which order of magnitude this time is roughly in" change from an unknown variable to a comparable option.

Thus, the configuration logic of large models on the enterprise side is the same. According to the complexity of post tasks, allocating low-multiplier models to high-frequency light tasks and high-multiplier models to low-frequency heavy tasks is one of the few levers that enterprises can actively operate under the premise of objective existing model price differences.

For Doubao, the significance of this design is not "accessing several more models", but that it begins to put the inference cost in front of users. In such paid products, letting users have an expectation of the cost of each task can maintain the willingness to pay more than simply reducing the price. The multiplier does not change the unit price, but changes users' judgment of "how long this money can last".

This is not only the layered management of the computing power bill, but also the use of "predictability" instead of "cheaper" to retain usage depth on the premise that the subscription price remains unchanged. In the final analysis, Doubao Work uses the multiplier to answer the questions that must be answered after the entrance expands: how to stratify the computing power bill, how to implement the commercialization signal, and how to calibrate the unit economic model.

After completing the admission ticket, the real test is delivery and renewal

Calculating the cost clearly is internal skill, and winning the choice is external battle.

For Doubao Work, it is entering a track that is expanding rapidly in scale but the head position has been initially solidified.

According to Analysys' "China Desktop AI Native Office Intelligent Agent Report" for the third quarter of 2026, the total monthly visits of 20 products included in the statistics in September exceeded 300 million, an increase of nearly 5 times compared with June 2026. Among them, Tencent WorkBuddy ranked first with 85.34 million monthly visits, Alibaba Qwen Office had 46.55 million visits, and ByteDance Doubao Work had 30.39 million visits. The total of the top three products exceeded 160 million visits. WorkBuddy's lead is fault-type, almost twice that of the second place.

It should be noted that the monthly visits of desktop AI native office intelligent agents counted by Analysys are not the same statistical object as the overall monthly active users of the Doubao App. After Tencent merged WorkBuddy, CodeBuddy and Marvis, its monthly visits reached about 117 million, taking the lead in entering the 100 million scale. The office Agent is far from the final outcome, but "who is leading" is no longer an open question.

WorkBuddy achieved this position not because Hunyuan is absolutely leading in every task. According to Sina Tech, when WorkBuddy was officially launched, the domestic version could seamlessly switch between Hunyuan, DeepSeek, GLM, Kimi and MiniMax.

36Kr's judgment in analyzing its growth path is more straightforward: WorkBuddy does not bind a single model, but connects multi-models, Skills, local tools and Tencent's ecosystem together, allowing users to freely choose the model with the highest cost performance and most suitable for the current task, reducing the migration cost to zero.

It caters to the most simple demand of office workers: "I don't care what parameter level model you are, I only care if you can make a chart that does not need rework in one minute". For WorkBuddy, multi-model was not a strategic action that needed to be announced from the very beginning, but the default configuration of the product.

However, the office Agent track does not have only one playing method. In the current domestic AI office track, Alibaba Qwen Office and Tencent WorkBuddy are both main players, and the three have obvious differences in ecological foundation and capability positioning.

From the perspective of ecological binding, Tencent Office and WorkBuddy are closely relying on the WeChat and WeCom ecosystem, with stronger adaptability in instant messaging and lightweight document collaboration scenarios, and more suitable for small and medium-sized enterprises that highly rely on the WeChat system.

Doubao Work is rooted in the Feishu ecosystem, with more prominent performance in project management, process promotion and complex workflow disassembly, and higher matching degree for large and medium-sized enterprises and knowledge-intensive teams.

From the perspective of core capabilities, Qwen Office focuses on content generation and document collaboration, and is good at standardized office outputs such as text polishing and PPT production; WorkBuddy is more like a lightweight task assistant, focusing on quick Q&A and information query within WeCom; Doubao Work has extended to desktop operation, multi-Agent collaboration and independent promotion of the whole process, with a wider capability boundary, and is more suitable for undertaking complex, long-cycle work tasks.

From the perspective of terminal interaction, all three support multi-terminal use, but Doubao Work's capabilities such as local GUI operation, persistent operation on cloud computer and mobile remote control have gone further in hardware interaction and task continuity.

This explains why Doubao Work's access to third-party models is first and foremost a competitive equalization action, not a differentiated advantage. In this track, the width of the model lineup is becoming an admission ticket. If Doubao Work only provides self-developed models, users will have a minus item of "not enough models" when choosing.

At present, the competitive focus of the office Agent track is shifting from "whose model is stronger" to "who can finish the work and make people feel at ease". Therefore, Doubao's access to Kimi, GLM and DeepSeek is to ensure that the competition will not be decided at this level, and then have the opportunity to bring the comparison back to the place where it may truly establish barriers.

Some industry analysis believes that there is a natural misalignment in office Agent: employees use it, and the work results belong to the company. The fact that individuals use it smoothly only proves that the product has crossed the usage threshold, not that the whole work has entered the enterprise production. The renewal period is the window to test the real competitiveness, and the delivery implementation capability directly determines the enterprise's willingness to pay and the retention level.

In this sense, Doubao Work's access to third-party models is not to use Kimi and DeepSeek to fight against WorkBuddy, but to pull the competition from "model lineup width" back to the depth of "Agent orchestration and ecosystem interconnection".

For Doubao Work, the truly non-replicable parts are Feishu permission inheritance, local and cloud dual execution, and the context connection from office to creation within ByteDance's ecosystem. WorkBuddy's barriers are also not in the model list, but in the workflow entrance composed of WeCom, WeChat and Tencent Docs.

Wandian Research believes that the model width determines the qualification to sit at the table. Doubao Work's access to third-party models is to complete the admission ticket, not to win the game. The final test standard is not the model list, but who can finish the work and make the enterprise renew the subscription. Behind this admission ticket, how to stratify the computing power bill and how to calibrate the unit economy are the longer exam questions for Doubao Work.

Conclusion: The larger the entrance, the clearer the account must be calculated

In the second half of 2026, the global AI industry encountered considerable challenges.

According to the Wall Street Journal, documents submitted by Anthropic on November 10 show that Anthropic plans to go public as soon as next week, with a target valuation of about USD 1.3 trillion, which is a considerable shrinkage compared with the expected USD 2 trillion.

To find the reason from the fundamentals, according to CLS, by the third quarter, Anthropic's revenue was USD 16.956 billion, and the quarter-on-quarter growth rate dropped from 144% in the second quarter to 47%; the computing power and infrastructure expenditure in the current quarter was about USD 9.9 billion, accounting for 58% of revenue, and the operating profit turned to a loss of USD 567 million from USD 974 million in the second quarter; the revenue brought by every 1 dollar of computing power input dropped from 2.08 dollars to