After observing the market for several days, I find that most interpretations of Doubao's paid version are, unfortunately, incorrect.

In summary, the market's narrative around Doubao retaining a free version while introducing paid features for advanced functions largely falls into two categories:

The first is the "free trumps paid" narrative from the internet era:

In this narrative, Doubao, now with the second-highest monthly active users globally and first in China, will see its active users migrate to free models once a paid version is introduced, leading to a decline in DAU and affecting its valuation.

The second is the resurgence of painful memories from SaaS investors about the "difficulty of charging":

In this narrative, the Chinese market's limited software payment habits make the revenue potential of a charging model highly questionable—a drop in the bucket that won't make a difference.

In my view, both interpretations represent a misjudgment of AI development and reflect how the internet Public opinion circle 's (public opinion circle's) ability to judge the cutting edge of AI development has yet to catch up with the evolution of China's AI technical capabilities.

From our perspective, using Doubao as a case study, retaining free basic features while commercializing through a subscription model is an inevitable return to AI's tool-like nature as a value provider. More importantly, this process will enhance the product experience for both free and paying users.

However, current public opinion, constrained by classical internet-era traffic narratives or SaaS narratives limited by revenue logic, has become quite disconnected from the development logic of today's AI era.

01

Model Inference Has Real Costs: A Shared Challenge Among Leading AI Players

Unlike the "zero marginal cost" logic of the internet and SaaS, AI generates significant real costs daily, whether in training or inference.

Information about ByteDance's computational costs is limited, but let's make a rough annualized estimate using publicly available data:

According to media and Volcano Engine disclosures, as of March 2026, Doubao's daily Token consumption has reached 120 trillion, a 1,000-fold increase from its launch in May 2024 (when it was just 0.12 trillion per day). This growth reflects user demand far exceeding the early chatbot era, driven by the explosion of agents, multimodal tasks, and long-chain workflows.

ByteDance has demonstrated excellent control over computational costs:

For example, in its previous Pro version, ByteDance announced pricing as low as 0.0008 RMB per 1,000 Tokens, implying its self-built computational costs are even lower. However, even at this significantly reduced cost, Doubao's computational consumption remains staggering:

Based on Pro version pricing, Doubao's daily computational costs could reach up to 100 million RMB; even assuming computational costs are just 30% of the pricing, daily costs would still be in the tens of millions of RMB. This means Doubao's annualized inference hard costs easily exceed 10 billion RMB.

This doesn't even account for the massive investments in new model training, upstream capital expenditures, computational upgrades, and additional consumption from growing user numbers and long-task demands—each pointing to annual costs in the hundreds of billions.

Doubao's usage costs, starting at billions of RMB annually and growing exponentially, are worlds apart from the "zero marginal cost" logic of the classical internet era.

Annual computational costs for long tasks, easily reaching hundreds of billions and potentially trillions, make it nearly unsustainable for any internet giant to subsidize AI with advertising revenue alone. They will soon hit a cash burn wall and be forced to make choices:

Either significantly raise prices/impose limits or tap into the productivity potential of professional paying users.

02

Doubao's Pricing Logic Aligns with Leading Players Like OpenAI

Let's examine how other global AI leaders approach this:

From late 2025 to early 2026, OpenAI significantly enhanced Codex capabilities (code generation + Agentic Coding) but actively restricted free user access and slowed overall user growth.

Specifically, OpenAI tied its most powerful inference models (o1, o1-pro, Codex enhanced) deeply to ChatGPT Plus/Pro/Team paid plans. Free users could still use basic GPT-4o, but complex long tasks (multi-file refactoring, prolonged agent debugging, understanding million-line codebases) were heavily restricted or queued.

This directly led to a noticeable slowdown in OpenAI's weekly active user growth from late 2025 to early 2026—from tens of millions of new users per month to gradually stabilizing growth. But this was a deliberate move by OpenAI: prioritize serving professional developers and enterprises willing to pay, leading to a rapid rise in API and enterprise subscription revenue. Fiscal year revenue growth now comes primarily from APIs and enterprise subscriptions, not free consumer users.

In early 2026, OpenAI formally opened Codex agent capabilities to ChatGPT Pro, Team, and Enterprise users, explicitly stating initial "generous access" would later face usage frequency limits, requiring free users to purchase additional credits.

This is a classic example of "enhancing high-consumption capabilities + paid prioritization + free throttling," highly congruent with Doubao's current strategy.

As paid-enhanced ChatGPT capabilities improve, free users benefit too—this is the correct commercial flywheel.

It's clear that ByteDance and OpenAI face congruent situations. Doubao's position in China mirrors ChatGPT's globally; user growth is no longer the top priority—optimizing user experience through pricing differentiation matters more.

In fact, the closer an AI organization sets its North Star metric to "red envelopes, monthly actives, strong operations," the further its model capabilities drift from the holy grail of AGI. Models become trapped in shallow usage and traffic optimization, endlessly circling non-core metrics unrelated to productivity gains, wasting already expensive computational resources and organizational patience.

03

AI Is a Tool: Paying Users Provide Effective Feedback

Today, much AI discourse falls into a trap: AI subscriptions aren't about selling features but selling Token access rights under varying intensity and cost curves.

Productivity tools inherently involve costs; they essentially use pricing mechanisms to filter high-quality feedback sources, driving models toward greater practicality and reliability.

First, tool iteration requires clear reference points. While free models quickly scale user numbers, they often generate low-quality feedback. Free users are typically casual trialists with limited usage scenarios, providing superficial emotional or random test feedback—noise that masks genuine core needs.

In contrast, paying users directly invest in performance, response speed, stability, and context length. Their payment behavior signals genuine reliance and deep usage intent. This feedback holds greater commercial authenticity and improvement value, guiding teams toward optimal directions.

This mechanism has long proven effective in traditional software and SaaS:

For example, after Figma launched Professional and Organization paid plans, feedback from paying design teams (especially large enterprises) on high-frequency, complex collaboration scenarios directly drove rapid iteration of core features like multi-version history and AI-assisted design (generating variants, automatic layout optimization). These features far exceeded contributions from free users' shallow usage, ultimately establishing Figma's unassailable moat in design tools.

The same logic applies to Doubao: by setting a payment threshold, it locks in enterprise users, professionals, and deep adopters with real business scenarios. Through pricing, these productivity users avoid queue times caused by computational scarcity.

This means resource allocation prioritizes paying users: faster responses, no queuing, more stable complex task handling. Paid versions often access higher-performance or full-power models, reducing hallucinations, improving context retention, and maintaining smoothness in multi-round dialogues or heavy productivity tasks (e.g., PPT generation, long document analysis, high-def image generation).

Free versions impose daily limits on high-consumption features (e.g., advanced generation); once exhausted, usage is restricted or slowed. Paid versions not only raise limits substantially but add prioritization, ensuring heavy users have "on-demand" access when needed.

More importantly, amid global computational scarcity—where leading players like OpenAI, Anthropic, and Google all face shortages—Doubao's design aligns with tool-based resource allocation logic.

For casual users, daily light usage on the free version suffices without noticeable impact. For professionals, creators, and students with frequent needs, 68 RMB/month for "no lag, no waiting" efficiency gains is often worthwhile—directly translating to time savings and output improvements.

By diverting high-consumption users through paid tiers, overall computational resources are allocated more rationally, significantly reducing queue times for free users. This creates a win-win: "paid users get upgraded experiences + free users see improved experiences." This is the optimal resource allocation logic for tool-based products in scarce environments.

Simply put, tools inherently have costs; tool iteration needs clear reference points. Only paying users, willing to pay for performance, speed, and stability, provide feedback with commercial authenticity and improvement value. Free users' noise often masks real needs; payment behavior itself is the most precise signal-filtering mechanism.

04

AI's Destination Is AGI, Not Another Traffic Platform or SaaS

However optimistic we are, today's AI remains in its early stages of species proliferation. The long road to AGI requires deep feedback from paying users to truly drive models toward practical tools and AGI.

Along this path, revenue and multiples from enterprise and individual payments aren't critical. Using classical SaaS valuation multiples often leads to significant misjudgments. What matters is the model's progress in logical reasoning, knowledge application, and task execution. Paying users' real-world feedback is the most reliable input for this progress.

In their work, creation, and decision-making, paying users fully expose the model's hallucinations, context retention weaknesses, and domain knowledge gaps. This feedback directly informs iteration priorities, helping models transition from general chat tools to trustworthy productivity assistants.

Revenue growth in this process also becomes an important metric for productivity levels. Expanding paying user bases means the model's marginal value creation is market-recognized, providing Chinese AI companies with sustainable cash flow for further R&D investment, computational procurement, and data optimization—forming a virtuous cycle.

Conversely, relying solely on free models may yield short-term traffic but risks resource dispersion and feedback distortion, hindering long-term technological progress.

Anthropic, now a hot company valued at nearly $1 trillion, moved earlier and further on paid tiers:

In early 2026, Anthropic gradually removed or limited high-consumption Claude Code features from its basic Pro plan, pushing heavy users to upgrade to higher tiers or switch to API payments.

This directly helped Anthropic's inference gross margin surge from 38% to over 70%: Claude Code's weekly active users doubled within two months, with ARR (annualized revenue) rapidly climbing to tens of billions of dollars.

High-intensity feedback from paying users is the core driver of Anthropic's leading Agentic Coding capabilities.

Doubao's pricing strategy reflects long-termism: avoiding traffic anxiety, it accumulates high-quality signals through value exchange, steadily advancing AGI-related capabilities.

In fact, more Chinese large model companies should actively encourage users to pay based on model capabilities, using a superior business model to generate positive feedback for model optimization.

Bluntly put, lacking confidence to charge may indicate organizational instability.

05

Conclusion

In my view, Doubao's pricing move isn't too early—it's long overdue.

It reflects Doubao's proper awareness as China's large model pioneer: not obsessing over short-term leads in free traffic and user numbers but actively building a sustainable, healthy commercial closed loop (closed loop).

This step isn't just about self-sufficiency—it sets the right industry benchmark: AI's future belongs to models daring to be accountable to value payers. This is the only path to AGI.