Author | Hao Xin

Editor | Wu Xianzhi

Doubao, which boasts 345 million monthly active users and has long followed a 'mass-market' strategy, has recently made a sudden 'pivot.'

In early May, Doubao officially announced the launch of a paid subscription model, setting off a wave of discussions.

The paid subscription is divided into three tiers: Standard at RMB 68/month, Enhanced at RMB 200/month, and Professional at RMB 500/month. It primarily targets complex task scenarios such as PPT generation and data analysis, while basic functions remain free.

According to sources close to Doubao who spoke with Photon Planet, the paid model will initially undergo grayscale testing, with some users seeing membership rules in advance. 'If all goes as expected, it will be fully implemented by the end of this month.'

However, many have reported to us that after upgrading to the latest version of Doubao, they have not yet encountered the paid model, suggesting that large-scale grayscale testing may not have begun yet.

The source indicated that the Doubao team had indeed received user feedback regarding needs for AI-powered PPT creation, AI spreadsheets, and other features, which was one of the motivations for considering a paid model.

Doubao is not an early adopter of paid models. Internationally, OpenAI and Claude implemented subscription models from the outset, with prices rising as intelligence levels improved. Domestically, startups like MiniMax and Yuezhi'anmian have partially abandoned casual users in favor of productivity scenarios to meet listing and commercialization goals.

With this charge (paid) move, Doubao has not entirely abandoned its 'mass-market' approach but has chosen a more moderate 'middle path.'

Another insider told us that computing resources are the most critical consideration. Some professional scenarios, such as expert modes, PPT generation, and spreadsheet creation, are highly resource-intensive.

'Currently, there are no additional resource allocations, so many users encounter errors or 'try again later' prompts during use. With a paid model, we can ensure a smooth experience without such issues.'

Doubao's Gradualism

Various signs indicate that Doubao's previous free model has encountered resource constraints. While casual users experience no issues, the experience for heavy professional users has declined.

Internally, the Doubao team believes that the time is right for charging, based on two main points: First, there is genuine user feedback, with some users voluntarily expressing a willingness to pay for a stable experience—this is not an imagined demand. Second, the usage scenarios for AI assistants have clearly stratified.

The free version can cover most daily conversations and simple Q&A, while features like AI-powered PPT creation and intelligent spreadsheet editing are resource-intensive, have relatively low usage frequency, but offer strong perceived value. These low-frequency yet high-cost features are naturally suited for a subscription model.

Currently, the professional features Doubao offers are similar to those of other products on the market. This reveals that the primary selling point of the current paid model is not exclusive features but stable service quality. There is a sense of 'catching up' in terms of product maturity—ensuring usability before gradually differentiating features.

Thus, the most direct purpose of the paid subscription is not to exclude free users but to create a 'fast lane' for heavy users, using pricing to filter demand and ensure service stability. In summary: keeping basic functions free for the masses while guaranteeing performance for professionals.

The most positive aspect of Doubao's strategy is that it has not alienated free users. The official response emphasizes that 'basic functions remain free,' and the Doubao team has previously stated that 'most needs of casual users are covered by the free version.'

In contrast, Kimi's pricing strategy is more aggressive. To some extent, Kimi has abandoned traffic and casual users—even for simple questions, users must wait or are told they are in a 'peak computing period' and are encouraged to subscribe.

Doubao appears more restrained, aiming to maintain traffic while exploring commercialization. Its 'gradualist' approach is essentially a commercialization strategy that uses user scale as its foundation, computing costs as its boundary, and scenario stratification as its lever. This aligns with ByteDance's product philosophy of prioritizing scale, then monetization, while cautiously preserving the core experience.

ByteDance's massive traffic base does not allow for 'shock therapy.' Doubao has 345 million monthly active users, a strategic asset for ByteDance. If a full subscription model were suddenly implemented, significant user loss would occur, and competitors would quickly capitalize. The 'gradualist' approach maintains the free baseline, acting as a safety net to protect traffic while testing the ceiling.

If prices were high and forcefully pushed, users would undoubtedly feel it is not worth it. A more moderate compromise allows charging based on service quality first, with room to gradually raise prices or add tiers later as features improve.

At this stage, this approach is rational. It avoids massive user loss, gives the team time to develop differentiated features, and generates revenue to fund computing expansion.

According to sources close to ByteDance, following its usual operational logic, Volcano Engine uses token consumption as a core business metric, and the Doubao team previously included traffic growth in its performance evaluations. Once Doubao's paid model is fully implemented, it is highly likely that paid conversion rates will be added to the performance metrics later.

Computing Power, Still Computing Power

Extending the timeline, multiple ByteDance AI products began charging or adjusting prices in the first half of this year. Doubao's charge (paid) move has sparked controversy simply because C-end users are the most sensitive and easily triggered.

ByteDance's AI coding tool, Trae, adjusted its pricing scheme from 'per-use' to 'per-token' and introduced five subscription packages. The AI video generation tool, Jimeng, raised prices three times in a single month, with annual fees for premium memberships increasing and effectively devaluing points (points). Volcano Ark's Coding Plan package suspended its first-purchase discount in March and now offers two paid tiers without additional temporary discounts.

Doubao's charge (paid) move has ignited public debate, with some saying, 'Even cash-rich ByteDance is running out of money.'

Undeniably, computing resources are currently the biggest bottleneck affecting AI development. GPU lease (leasing) prices remain high, and domestic AI applications lack the USD pricing power enjoyed by OpenAI. As a result, all major model vendors subsidize AI with profits from other businesses.

Alibaba has e-commerce and cloud, ByteDance has advertising and Douyin, and Tencent has gaming. However, compared to Alibaba Cloud, which already has a vast enterprise client base and mature cloud service revenue, ByteDance's Volcano Engine started later and has fewer client resources, putting it at a disadvantage in monetizing computing leasing.

Thus, ByteDance needs two things more urgently than Alibaba: direct C-end charging and point-to-point GPU leasing for major clients. This explains why Doubao is charging—not because the product is perfect, but because the computing bills can no longer wait. Only professional scenarios justify user payment to cover the high computing costs per request.

There are reports that ByteDance is also investing in self-developed GPUs, which could be a future game-changer. If successful and deployed at scale, ByteDance's computing costs could drop significantly. At that point, Doubao might face two choices:

1. Lower prices to attract a larger pay (paying) user base and squeeze competitors' space.2. Maintain prices to increase profits while using saved computing resources to optimize the free version's experience, further widening the gap with rivals.

Besides computing factors, a clear signal from relevant sources is that ByteDance and Doubao believe 'China has reached a turning point in AI commercialization, with mass-market and professional tracks now diverging.'

ByteDance is building a closed-loop commercial ecosystem for its AI business. P-end and enterprise clients pay for the capabilities of products like Doubao, Trae, and Jimeng/Dreamina. Commercialization is a step toward actively optimizing the product ecosystem, achieving user stratification, and establishing a healthy commercial model for long-term AI infrastructure investment.

This is not just ByteDance's choice but also signifies that China's AI industry is establishing a commercialization paradigm. The era of maintaining free services solely through burning money is ending, and vendors are strategically seeking paid scenarios.

AI price hikes have surged this year. OpenAI's ChatGPT Plus rose from $20 to $30 per month, Claude Code was removed from the $20/month Pro plan, and Cursor Pro increased from $20 to $40. Divided into professional, premium, and top-tier memberships, Doubao's announced prices are significantly lower than those of ChatGPT, Claude, and Gemini abroad and slightly lower than Kimi's domestic pricing.

This is an inevitable result of technological maturity and the agent boom, essentially indicating that free models can no longer support the unlimited growth of high-load professional demands, necessitating tiered services.

The Third Path and Industry Bellwether

The market currently falls into two main camps:

The free-leaning camp, such as Qianwen, Yuanbao, and DeepSeek, backed by Alibaba, Tencent, and DeepSeek, respectively, have their computing costs covered by other group businesses or investors. Their primary goals are to acquire users, accumulate data, and drive traffic to cloud services. The sustainability of this model depends on the willingness of the financiers—once the group cuts budgets, free services may be compromised.

The other camp leans toward full commercialization, such as Kimi and MiniMax. Startups lack unlimited resources and must achieve self-sufficiency quickly. Thus, their free versions offer limited functionality, forcing users to either pay or endure lower service quality or usage limits. This model can survive but struggles to grow user scale and risks ceding free users to competitors.

Doubao is now attempting a third path: keeping basic functions free to maintain traffic while charging for professional scenarios to cover high computing costs.

Regardless of how Doubao's commercialization exploration progresses, it holds significant implications as a bellwether for the industry.

In reality, the paid conversion rate for AI assistants in the industry is typically very low. If Doubao can achieve a 1-2% paid conversion rate among its 345 million monthly active users—i.e., approximately 3.5-7 million paying users—it would demonstrate a replicable commercialization model for domestic AI applications. Products with hundreds of millions of monthly active users could sustain operations with just 1% core users, greatly boosting capital confidence in the AI application layer.

More importantly, Doubao's commercialization model does not require users to change their habits or vendors to abandon scale. The resources and ecosystems built by major players can instead become their unique strengths.

If this template proves successful with 345 million users, it will become the easiest path for most AI assistant products to emulate, supporting the future valuation of an AI application company.

Doubao's 'third path' resembles a test of the industry's payment threshold.

Doubao's value lies not in how much money it makes but in showing the industry whether, under current computing costs, an AI assistant with hundreds of millions of users can achieve sustainable operations through tiered services. It reveals what capabilities users are willing to pay for, how much, and where the boundary between free and paid lies.

These answers are needed not just by ByteDance but also by Alibaba, Tencent, Baidu, and all AI startups.