After observing the market's reactions for several days, I've found that most interpretations of Doubao's subscription model are unfortunately misguided.

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

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

In this narrative, Doubao, now ranked second globally and first in China for monthly active users, risks losing active users to free models if it introduces a paid tier. Declining DAU would undermine 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 willingness to pay for software makes the revenue potential of a subscription model highly questionable—a drop in the bucket.

From my perspective, both interpretations represent a misjudgment of AI development and reflect the internet Public opinion circle 's (public opinion sphere's) failure to keep pace with China's evolving AI capabilities.

In our view, using Doubao as a case study, retaining free core features while commercializing through subscriptions is an inevitable return to AI's tool-like nature as a value-delivering technology. More importantly, this process will enhance the product experience for both free and paying users.

Current public discourse, constrained by classical internet-era traffic narratives or SaaS logic focused on revenue scale, has become disconnected from the development logic of today's AI era.

01

The Real Costs of Model Inference: Aligning with Leading AI Players

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

Limited information exists about ByteDance's computational costs, but we can make a rough annual estimate using publicly available data:

As of March 2026, Doubao's daily token consumption reached 120 trillion, a 1,000-fold increase from its launch in May 2024 (when it handled just 0.12 trillion tokens per day). This growth reflects user demand far exceeding early chatbot-era usage, driven by the explosion of agents, multimodal tasks, and long-chain workflows.

ByteDance has demonstrated exceptional cost control over computational expenses:

For example, its Pro version priced tokens as low as 0.0008 RMB per thousand, implying its self-built computational infrastructure costs even less. However, even at these significantly reduced rates, Doubao's computational consumption remains staggering:

Based on Pro version pricing, Doubao's daily computational costs could reach 100 million RMB; assuming costs are just 30% of pricing, daily expenses still likely amount to tens of millions of RMB. This translates to annualized inference costs easily exceeding 10 billion RMB.

This excludes expenses for 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 costs, starting at billions of RMB annually and growing exponentially, stand in stark contrast to the "zero marginal cost" logic of the classical internet era.

Annual computational expenses for long tasks, ranging from hundreds of billions to trillions, make it nearly impossible for internet giants to subsidize AI indefinitely through advertising revenue. They will soon hit a cash wall, forcing tough choices:

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

02

Doubao's Subscription Logic Mirrors Global Leaders 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) while actively restricting free user access and slowing overall user growth.

Specifically, OpenAI tied its most powerful inference models (o1, o1-pro, enhanced Codex) to ChatGPT Plus/Pro/Team paid plans. Free users retained access to basic GPT-4o but faced severe restrictions or queuing for complex long tasks (multi-file refactoring, prolonged agent debugging, understanding million-line codebases).

This directly caused OpenAI's weekly active user growth to slow markedly in late 2025–early 2026—from tens of millions of new users monthly to stagnant growth. But this was a deliberate strategy: prioritizing paying professional developers and enterprises, with API and enterprise subscription revenue rising rapidly. Fiscal year growth now stems primarily from APIs and enterprise subscriptions, not free consumer users.

In early 2026, OpenAI formally opened Codex agent features to ChatGPT Pro, Team, and Enterprise users, initially offering "generous access" before imposing usage frequency limits, requiring free users to purchase additional credits.

This follows the classic playbook of "enhancing high-consumption capabilities + prioritizing paid users + throttling free access," highly aligned with Doubao's current strategy.

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

Clearly, ByteDance and OpenAI face identical challenges. Doubao's market position in China rivals ChatGPT globally. User growth is no longer the top priority; optimizing user experience through pricing tiers matters more.

In fact, when AI organizations prioritize "red envelopes, monthly active users, and intense operations" as their North Star metrics, their model capabilities drift further from AGI—the ultimate goal. Models become trapped in superficial usage and traffic optimization, endlessly circling non-core metrics unrelated to productivity gains, wasting already expensive computational resources and organizational patience.

03

AI as a Tool: Paying Users Provide Effective Feedback

A common misconception in today's AI discourse is that AI subscriptions sell features. In reality, they sell token access rights under varying intensity and cost curves.

Productivity tools inherently incur costs. Their purpose is to filter high-quality feedback sources through pricing, driving models toward practicality and reliability.

First, tool-based product iteration requires clear benchmarks. While free models rapidly scale user numbers, they often generate low-quality feedback. Free users, typically casual experimenters with limited use cases, provide superficial or random test feedback—noise that obscures genuine core needs.

In contrast, paying users directly invest in performance, response speed, stability, and context length. Their willingness to pay signals genuine dependency and deep engagement. This feedback holds commercial authenticity and improvement value, guiding teams toward optimal refinements.

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

For example, after Figma introduced Professional and Organization paid plans, feedback from paying design teams (especially large enterprises) on high-frequency, complex collaboration scenarios directly accelerated iterations of core features like multi-version history and AI-assisted design (variant generation, auto-layout optimization). These features far exceeded contributions from free users' superficial usage, cementing Figma's dominance in design tools.

The same logic applies to Doubao: by setting payment barriers, it targets enterprise users, professionals, and power users with real business scenarios. Through subscriptions, these productivity-focused users avoid queuing delays caused by computational scarcity.

This means systems prioritize paid requests: faster responses, no queuing, and stable handling of complex tasks. Paid versions often access higher-performance or full-capacity 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-definition image generation).

Free versions impose daily quotas on high-consumption features (e.g., advanced generation), throttling or slowing access once exhausted; paid versions drastically increase quotas and add priority rights, ensuring power users "get what they need when they need it."

More importantly, amid global computational scarcity—a reality for OpenAI, Anthropic, Google, and others—paid tiering alleviates resource conflicts. Doubao's design aligns with tool-based resource allocation logic.

For casual users, the free version suffices for daily light use without noticeable impact. For professionals, creators, and students needing frequent access, 68 RMB/month for "lag-free, wait-free" efficiency is often worthwhile—directly translating to time savings and output gains.

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

Simply put, tools incur costs, and iteration requires clear benchmarks. Only paying users, willing to invest in performance, speed, and stability, provide feedback with commercial authenticity and improvement value. Free users' noise often masks true needs; payment itself is the most precise signaling mechanism.

04

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

However optimistic we may be, today's AI remains in its early stages of species-level proliferation. The long journey toward AGI—the ultimate prize—depends on deep feedback from paying users to drive models toward practical tools and AGI.

Along this path, revenue and multiples from enterprise and individual payments matter less than they seem. Classical SaaS valuation multiples often lead to significant misjudgments. What truly matters is progress in logical reasoning, knowledge application, and task execution. Feedback from paying users' real-world scenarios provides the most reliable input for this progress.

In their work, creativity, and decision-making, these users expose model hallucinations, context retention failures, and domain knowledge gaps. This feedback directly informs iteration priorities, helping models evolve from general chatbots into trustworthy productivity aids.

Revenue growth becomes a key productivity metric in this process. Expanding paying user bases signals market recognition of the model's marginal value creation, enabling sustainable cash flow for Chinese AI firms to reinvest in R&D, computational procurement, and data optimization—forming a virtuous cycle.

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

Anthropic, now valued at nearly $1 trillion, moved earlier and further on paid tiering:

In early 2026, Anthropic gradually removed or limited high-consumption Claude Code features from its base Pro plan, pushing power users to upgrade to higher tiers or API subscriptions.

This directly lifted Anthropic's inference gross margins 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 has been the core driver of Anthropic's leading Agentic Coding capabilities.

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

More Chinese large model firms should actively encourage users to pay based on model capabilities, creating superior business models that positively reinforce model optimization.

Frankly, a lack of confidence in charging may signal organizational instability.

05

Conclusion

In my view, Doubao's subscription move is not premature—it's overdue.

It demonstrates the maturity expected of China's pioneering large model player: eschewing short-term leadership in free traffic and user numbers, it proactively builds a sustainable, healthy commercial closed loop (closed loop).

This step not only generates revenue but also sets a correct industry benchmark—AI's future belongs to models accountable to those who pay for value. This is the only path to AGI.