Written by: Zhang Feng

In May 2026, ByteDance’s AI product Doubao revealed a paid subscription testing plan in the application store, officially commencing commercialization exploration targeting the consumer market, marking a significant event as the first mainstream large model application in China to move out of the "completely free" stage.

Against the backdrop of the industry largely relying on free traffic expansion and unclear commercialization paths, Doubao's paid testing has sparked widespread discussion: some view it as an inevitable choice for the large model industry to say goodbye to the cash-burning model and transition to self-sufficiency, while others question whether ordinary users are willing to pay continuously for an AI dialogue tool.

However, beyond the superficial pricing controversy and discussions about user acceptance, a deeper and more disruptive industry transformation is quietly taking place— the main user of AI is shifting from "humans" to "agents," with the core logic of human-computer interaction moving from "command execution" to "autonomously completing tasks." The true value of Doubao's paid exploration lies not in the amount of short-term membership revenue, but in its ability to seize the largest strategic opportunity window in the AI field for the next decade—the transition from a dialogue assistant to the vital infrastructure of an agent ecosystem.

1. Doubao's Paid Model Analysis: A Realistic Choice for Commercialization Exploration

1.1 Paid Model and Rights: Hierarchical Pricing Based on Productivity

Doubao currently adopts the industry-standard Freemium model of free basic functions plus paid value-added services. As of May 2026, the paid plan is still in the internal testing phase, with no official paid entry open in the product; final pricing and rights will be subject to official announcements. This test is aimed at users with high productivity needs, setting three subscription tiers, with core differences reflecting computational power consumption, professional functions, and service experience:

Free version: Retains core functions such as daily conversations, information retrieval, basic writing, and general Q&A, free for the general public, with advanced features having a daily free usage limit; response speed may be limited during peak periods.

Standard version: 68 yuan per month, 688 yuan per year, unlocking long document analysis, batch text processing, basic PPT generation, watermark-free export, and offering more stable response speed.

Enhanced version: 200 yuan per month, 2048 yuan per year, adding rights such as deep data analysis, multimodal content generation, and exclusive acceleration channels; suitable for professional creators and workplace users.

Professional version: 500 yuan per month, 5088 yuan per year, aimed at enterprise users and heavy users, providing full feature unlocking, exclusive customer service, priority computational resources, and customizable knowledge base as top-tier services.

From a product logic standpoint, Doubao's paid system is not merely about "unlocking functions," but is layered according to computational consumption and value scenarios. The free layer bears user outreach and basic service functions, while the paid layer focuses on high-cost, high-value productivity scenarios, attempting to guide users from "casual chatting" to "efficiency enhancement," which is also the mainstream direction for current large model commercialization targeting consumers. The official has not disclosed technical parameters such as context windows and token consumption; related descriptions are subject to public information.

1.2 Deep-seated Motives Behind the Paid Model: Cost Pressure and Breakthroughs in Business Models

Doubao's decision to advance the paid testing at this point is not a short-term tactical adjustment, but an inevitable result of the large model industry's development to the present stage, driven by a threefold pressure of costs, business models, and user values.

First, high computational costs make traffic scale a burden. Although AI chip technology iterations and model optimizations over the past year have significantly reduced inference costs, supporting tens of millions of daily active users with continuous conversations, document analysis, multimodal generation, and other services still requires investment of tens of millions in computational resources monthly. ByteDance's traffic advantage in the AI era has shown duality: the larger the user scale, the more computational consumption, and the more significant the loss pressure, making it difficult to maintain long-term with mere group financial support.

Second, the traditional advertising model has failed in the AI assistant scenario. Unlike information flow products such as Douyin and Toutiao, Doubao's core scenario is a one-on-one dialogue interaction actively initiated by users, lacking inherent advertising placements. Forcing ads into the conversation would directly undermine user experience, lower product usage willingness, and lead to advertising monetization efficiency far below that of traditional internet products. Moreover, model calling services targeting B-side also face fierce price competition from vendors like Baidu Wenxin, Alibaba Tongyi, and Deep Aigo, severely squeezing profit margins.

Third, there is a need for user value screening and product positioning upgrades. Current domestic C-end users still perceive AI assistants as "entertainment tools," with generally low willingness to pay. By setting up paid tiers, it can precisely filter out high-value users who have genuine productivity needs and are willing to pay for efficiency, while also pushing the product teams to focus on "solving real problems," rather than staying in meaningless chit-chat interactions, driving Doubao's transformation from a "general entertainment dialogue robot" to a "professional productivity tool."

2. Cognitive Reconstruction: Who Is the True User in the AI Era?

2.1 The Overlooked Core Premise: From "Humans Using AI" to "Agents Using AI"

Since ChatGPT ignited the global AI craze, the industry has almost unanimously accepted an unquestionable premise: the users of AI are humans. Product design revolves around the dialogue interface, the business model pursues "user scale x single-user revenue," and the charging model targets individual subscriptions, with all logic built on "humans directly operating AI."

This framework fits the product thinking of the internet era but is becoming a cognitive trap that hinders the industry from clearly seeing the future. We can envision a typical scenario in 2030: a user tells a smart device in the morning, "Help me plan my business trip for the week," after which no manual operation is required—AI automatically reads the calendar schedule, calls maps to plan the optimum route, queries and books flights and hotels, syncs travel information to the corporate expense system, and reminds about travel precautions. In this process, the user only completes the goal setting, while the true entities interacting with various systems, APIs, and data are the agents. This signifies that the direct users of AI are no longer humans, but agents capable of autonomously executing tasks, where humans are only responsible for proposing demands and accepting results.

This is not a distant science fiction imagination but a technical trend that is happening. Open-source projects like AutoGPT, BabyAGI, and CrewAI have already validated the feasibility of multi-agent collaboration; OpenAI's GPTs and Assistants API essentially build the underlying framework for agent development; AI products from technology giants like Google and Microsoft are accelerating the advancement of autonomous execution and cross-tool collaboration capabilities. Agents are transitioning from concept to reality, fundamentally changing the way AI is used.

2.2 Interaction Paradigm Revolution: From "Human in the Loop" to "Autonomous Execution"

Currently, the way humans interact with large models essentially replicates the industrial age assembly line model: humans issue clear commands, models execute and return results, and humans adjust commands based on results to proceed to the next step. This "human-in-the-loop" mode completely ties efficiency bottlenecks to human attention, decision speed, and operational ability, failing to leverage AI's automation advantages.

In contrast, the interaction logic of the agent era represents a fundamental disruption: humans only need to set clear goals, constraints, and priorities, while agents can autonomously complete the entire process of task decomposition, tool invocation, logic reasoning, exception handling, and result feedback. Multiple agents can also collaborate, negotiate, and divide tasks, forming a self-operating digital ecosystem. For example, a travel planning agent can link with flight agents, hotel agents, and itinerary agents to collectively create a one-stop trip plan; workplace agents can connect with document agents, data agents, and communication agents to automatically generate weekly reports, analyze business data, and synchronize work progress.

In this paradigm, humans transition from "direct operators" to "goal setters and supervisors," while AI shifts from "passively responsive tools" to "autonomous executors," leading to a logarithmic increase in efficiency, depth, and value of human-computer interaction.

2.3 Disruptive Reconstruction of Business Models

When agents become the true users of AI, the existing business logic based on "individual users" will become entirely invalid:

First, personal charging for C-end users neglects the real value consumption entity. When agents execute complex tasks, they generate massive computational consumption, tool invocation, and data interaction, which substantially exceeds the costs of personal daily conversations, rendering personal subscription fees unable to cover real costs.

Second, the pricing system based on accounts cannot adapt to the usage characteristics of agents. Individual users pay per account, whereas agents operate on a "task" basis, wherein a single task may involve multiple agents across various platforms, and account boundaries become entirely obsolete.

Third, billing models based on conversation counts and tokens mismatch severely with agent needs. An agent can complete thousands of API calls in a few seconds, consuming hundreds of thousands of tokens; traditional billing methods either lead to excessive user costs or result in significant losses for the platform.

This implies that Doubao's current paid testing still operates under the outdated assumption of "users are humans." Without timely adjustments to pricing logic, product form, and business models, it risks becoming passive in the face of the arrival of the agent era.

3. New Business Paradigm in the AI Era: From SaaS to AaaS

3.1 Core Leap: From SaaS to AaaS (Agent as a Service)

Over the past decade, SaaS (Software as a Service) has been the core business model of enterprise services and the digital economy, essentially reflecting "humans using software tools to complete work." Users pay for software usage rights, but data organization, process execution, and result output still require manual intervention, making tool value heavily dependent on human operational capabilities.

In the AI era, the business model is rapidly evolving from SaaS to AaaS (Agent as a Service). The essence of AaaS is not "humans controlling AI," but "agents replacing users to complete tasks and deliver results." The differences between the two are fundamentally distinct: SaaS sells "capability permissions," whereas AaaS sells "task results"; SaaS requires user active operation, whereas AaaS achieves autonomous execution; the value ceiling of SaaS is the tool itself, while the value ceiling of AaaS is the ability to solve problems.

For example, in market analysis scenarios: under traditional SaaS models, users need to purchase multiple tools for data analysis, visualization, and document editing, manually collect data, create charts, and write reports; whereas in AaaS models, users merely state their needs to the market analysis agent, which automatically scrapes publicly available data, invokes large models for analysis, generates professional charts, formats, and outputs full reports, with users receiving usable results without any manual operation involved. This "delivering results" model aligns more closely with core user needs, and the commercial value far exceeds that of traditional tools.

3.2 Three-layer Business Architecture Centered on Agents

The future AI industry’s business system will build a clear three-layer architecture around agents, with each layer having distinct positioning and business models:

Lower Layer: Agent Development Platform. As the infrastructure of the AI ecosystem, it provides core capabilities such as agent development frameworks, runtime environments, tool invocation, memory management, and knowledge bases, lowering the barriers for agent development. Typical representatives include OpenAI Assistants API, ByteDance Coze, Dify, etc., with business models primarily revolving around API invocation fees, computational resource fees, and developer seat fees.

Middle Layer: Agent Application Market. This serves as the "app store" of the AI era, focusing on agent distribution and matching. The platform gathers specialized agents in vertical fields such as travel planning, legal review, code development, workplace office, etc., allowing users to select and customize according to their needs. Business models include transaction commissions, subscription sharing, and traffic recommendation fees, with the core goal connecting developers and end users.

Upper Layer: Agent as a Service. This is the service layer directly facing end-users, hiding underlying technological details; users need not concern themselves with models, APIs, or tools, focusing solely on "whether the task is completed." Business models revolve around performance-based payments, such as commission charged upon successful ticket booking, fixed fees upon tax completion, and profit-sharing based on compensation amounts, creating a binding of interests between users and the platform.

In light of this architecture, Doubao's current positioning is quite awkward: it is neither an open agent development platform nor has it established a mature application market ecosystem, nor does it possess autonomous execution and result delivery capabilities typical of AaaS. It fundamentally remains at the stage of "enhanced dialogue robot," facing the risk of being eliminated in the agent era.

4. Charging Models for Agents: Breaking the Traditional Pricing Dilemma

4.1 Structural Conflicts in Traditional Pricing Systems and Agent Needs

Doubao's current paid model faces three irreconcilable conflicts with the agent era’s demands:

Conflict in Pricing Units: Individual users are sensitive to conversation counts and token numbers, whereas agents operate with high frequency calls and batch execution, making traditional pricing units completely unsuitable.

Conflict in Value Anchoring: Individual users pay for "convenience, information," resulting in low value; agents pay for "task completion, result delivery," which can enhance value by one or two orders of magnitude.

Conflict in Payment Entities: Traditional payments focus on individuals, while in agent scenarios, multi-agent collaboration and delegated execution lead to longer payment chains and more complex settlements, necessitating a new clearing and settlement system.

These conflicts determine that the functionality paid model based on individual subscriptions is destined not to become the mainstream business model in the agent era.

4.2 Four Viable Charging Models Adapted for Agents

As agent technology matures, the industry is exploring charging methods more in line with its essence. The following four models are feasible for implementation:

Task Completion Charges (Payment upon Results). This is the model that most aligns with the core value of agents, where users pay not for the process but for the successfully delivered result. For example, an insurance claims agent would take a commission after successfully recovering compensation, and a job application agent would charge a fee after helping to secure an offer, tying the interests of the platform and users closely, though with extremely high demands on agent capabilities.

Resource Reservation Charges. Drawing from cloud computing models, users or enterprises pre-purchase computational power, storage, tool invocation quotas, and other "resource pools," dynamically consumed by agents executing tasks. This is suitable for high-frequency usage scenarios, cost-controlled, and stable, making it the first choice for enterprise users.

Agent Collaboration Revenue Sharing. Complex tasks completed by multiple agents collaboratively, with revenue automatically distributed based on each agent's contribution. For instance, in a travel planning task, the itinerary agent, flight agent, and hotel agent share revenue according to an agreed proportion, constructing a "machine economy" micro-payment system.

Micro Subscription Aggregation. Users may need to use dozens of vertical agents, and individually subscribing can be too costly. A unified subscription aggregator allows users to pay a monthly fee, with the backend automatically settled and divided according to usage, simplifying the payment process and enhancing user experience.

4.3 Doubao's Payment Dilemma: Double Limitations in Thinking and Capabilities

From the agent perspective, Doubao's current paid testing presents two core issues:

Firstly, the pricing mindset remains focused on personal dialogue scenarios. The monthly subscription fee of 68 yuan essentially sells "a better conversation experience, more functional counts", rather than “task completion and value delivery”, failing to match the high-value demands of the agent era.

Secondly, the product capabilities do not support autonomous operations for agents. Doubao has not fully opened its API for external agent calls, has not provided a comprehensive agent development framework, nor has it integrated a sufficiently large tool ecosystem, only enabling passive dialogue and failing to autonomously execute complex tasks across platforms and tools, making it difficult to support the AaaS model's implementation.

5. The Emergence of Agent Applications: Technological Turning Points and Doubao’s Potential Advantages

5.1 From "Dialogue" to "Action": The Core Leap in Agent Commercialization

Currently, all large model applications face a consensus in the industry: dialogues are cheap, while actions and results are valuable. Users are willing to pay for agents that save money, earn money, or save time, but it is challenging to get them to pay long-term for purely dialogue robots. To truly activate the market, agents must complete three key leaps:

From passive responses to proactive execution: Agents will no longer wait for user instructions but will act actively based on changing scenarios, such as automatically adjusting itineraries and reordering services upon detecting flight delays.

From single-turn dialogues to multi-step reasoning: Agents will possess the ability to autonomously decompose complex tasks, formulate execution plans, and dynamically adjust strategies to handle uncertainties in real-world scenarios.

From information provision to result delivery: The ultimate goal is to "get things done" rather than merely informing methods, directly outputting actionable, usable results to achieve leaps in value amplitude.

5.2 Approaching Technological Turning Points: An Explosive Period in 12-18 Months

The core technical barriers obstructing the large-scale application of agents are rapidly being overcome:

Long context capabilities: The context windows of large models are expanding continuously, capable of carrying complete task histories and complex instructions, supporting agents' memory.

Tool invocation capabilities: From function calls to multimodal invocation protocols, large models' abilities to connect with external systems and invoke third-party tools are maturing.

Autonomous planning capabilities: Iterations of prompt engineering like ReAct, thinking chains, and thinking trees, as well as the deployment of dedicated planning model training, bring agent decision-making abilities close to practical levels.

Long-term memory mechanisms: The maturity of vector databases and memory storage technologies allows agents to accumulate experiences across tasks and sessions, continuously optimizing execution results.

The arrival of this technological turning point signifies that the next 12-18 months will become an explosive window for agent applications, and players who construct their ecosystems early will dominate the industry landscape in the next phase.

5.3 Doubao's Unique Advantage: Strategic Support from the ByteDance Ecosystem

Although its current positioning is lagging, Doubao is not without opportunities, as ByteDance possesses rare resources for building an agent ecosystem:

Traffic entry advantage: Doubao itself has a massive user base, with ByteDance products like Douyin, Toutiao, and Feishu potentially serving as natural channels to reach agents, quickly covering the general public and workplace users.

Scenario coverage advantage: High-frequency scenarios like content creation, e-commerce transactions, local living, and workplace office are all prime soil for agent implementation, capable of quickly validating commercialization value.

Technical and computational support: ByteDance has a solid foundation in the underlying technologies and application development layers of large models, with the Volcano Engine providing ample computational resources to ensure stable agent operation.

Coze platform: As ByteDance's one-stop agent development platform, Coze supports low-code/no-code construction of agents, multi-channel publishing to Doubao, and invocation of rich plugin tools, serving as a core vehicle for building the agent ecosystem.

If these resources can be effectively integrated, Doubao fully has the potential to transform from a dialogue assistant into an agent operating system and application store, becoming a core infrastructure in the AI era.

6. Four Major Risk Barriers to Agent Scaling

Despite the broad prospects, agents must still overcome four major layers of risks to mature commercially:

Technical risk: Agent loss of control and unpredictability. Once agents gain operational permissions, they may misoperate due to instruction misunderstandings or logical loopholes, such as a shopping agent making erroneous orders or a code agent causing system crashes. Multi-agent collaboration may also produce "emergent behaviors," with overall system behavior unpredictable, posing very high risks in critical scenarios like finance and industry.

Security risk: Challenges in identity and permission management. The traditional "user-account-permission" model is no longer applicable; one user can delegate multiple agents to execute tasks, leading to significant differences in permission scopes. How to prevent agents from stealing identities, how to set minimal permissions, and how to ensure data security lacks mature solutions.

Economic risk: The dilemma of value distribution. After multiple agents collaboratively complete tasks, how are the revenues fairly distributed? Conflicts of interest may arise among parties, while the rapid negotiation speed among agents makes it difficult for humans to supervise, necessitating a transparent and tamper-proof automated distribution mechanism.

Regulatory risk: Legal responsibility blanks. When agents cause losses, there are no clear legal provisions for attribution of responsibility: is it the developer, the platform, the user, or collaborating partners? Applications in sensitive fields like healthcare, transportation, and finance may face scaling difficulties due to ambiguous liability.

These risks require collaborative solutions from technology, industry, and regulatory bodies, and will also affect the speed and boundaries of agent commercialization.

7. Doubao's Future Depends on Agent Strategic Choices

Returning to the core question: How far can Doubao's paid exploration go? The answer lies not in the height of subscription prices, but rather in whether it can seize the strategic opportunity of the agent era.

If Doubao continues to adhere to the "dialogue assistant + function subscription" route, even if it obtains some membership revenue in the short term, it will eventually be eliminated in the agent era—a product that can only engage in dialogue, cannot act, and fails to build an ecosystem is destined to become a "transient" in the industry's evolution. However, if Doubao can complete three key strategic transformations, it will thoroughly open up growth spaces:

First, transform from a single dialogue product into an open agent platform. Fully open APIs, provide a comprehensive development framework, introduce third-party agents, and build an application distribution market, evolving from a "single application" to an "ecosystem vehicle."

Second, transform from charging based on features to charging based on results. Abandon the purely membership fee model, focus on task completion and value delivery, and implement commission-based and performance-based payment systems to unlock commercial value potential.

Third, shift from serving individual users to servicing the agent ecosystem. Expand core users from individuals to all agents on the platform, providing infrastructure services such as computational power, memory, permissions, and settlements, charging developers and enterprises.

ByteDance has already assembled key components like the Coze platform, Doubao entrance, Volcano Engine computational resources, and an all-scenario ecosystem; it only lacks a unified top-level strategy to integrate them into a complete agent ecosystem. The true significance of Doubao's current paid testing lies not in immediate monetization but in compelling the industry and companies to rethink: in the AI era, should the core positioning be to create a trend-following dialogue product, or to become the infrastructure defining the agent economy?

The opportunity window for agents has opened and will not wait for long. The next 12-18 months represent a golden period for building ecosystems, setting rules, and occupying minds. Whether Doubao can grasp this opportunity and complete the leap from paid testing to ecological transformation will determine its own commercial fate and influence the competitive landscape for China's AI industry in the global agent era. The ultimate answer to AI commercialization lies not in dialogues, but in the autonomous actions of agents.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。