Doubao has begun charging its users, with an annual subscription fee reaching up to 5,088 yuan, putting its 345 million monthly active users through a financial stress test.

In May 2024, Doubao burst onto the scene, captivating the market with a seemingly "suicidal" pricing strategy of just 0.8 yuan per million tokens. At that time, the focus wasn't on profit but on lowering barriers to entry to attract a vast number of individual and corporate clients. This move secured Doubao a first-mover advantage in the nascent market and plunged domestic large-scale models into a fierce price war.

By 2026, with internet barriers fortified and domestic access to overseas large models further restricted, coupled with the release of the revolutionary Seedance 2.0 product, Doubao, which had long championed a free model, took the lead in introducing a "three-tier paid plan" for users, with a maximum monthly fee of 500 yuan.

Within the common narrative advocating for "Doubao to start charging," several recurring arguments emerge.

The most frequently cited is that the marginal costs of AI are not zero—a concept that, while technical, is widely understood.

For products like Toutiao and WeChat, more users are akin to streetlights illuminating every corner, incurring essentially no additional costs with a high user base. However, every new user served by an AI large model consumes computational power, resources, and time, with no cost reduction from economies of scale.

Combined with ByteDance's net profit falling by more than 70% year-on-year in 2025 and internal concerns that "Doubao's commercialization path is unclear, and high DAU inference costs are severely squeezing profits," it's evident that ByteDance is under financial strain.

The second point is that Doubao needs to refine its user base.

Doubao boasts 345 million monthly active users, with the highest penetration rate, yet the majority are students and middle-aged and elderly groups.

For a long time, Doubao has been perceived as "simple-minded, prone to errors, but emotionally fulfilling." It's suitable for writing and chatting, but when it comes to professional fields, high-end users generally prefer ChatGPT, Claude, or some domestic vertical models.

ByteDance's paid professional version aims to showcase the robust capabilities developed over the past two years.

Paid users are not only heavy token consumers but also valuable data contributors, providing data and training models for the platform while paying for the privilege.

The third point is aligning quality with price—only by paying can high-quality services be ensured.

Transitioning AI from flashy demonstrations to value-based pricing is acceptable, but the services provided must distinctly feel superior, akin to the difference between first-class and economy class on an airplane.

Currently, large models have similar capabilities. Honestly, if you're just looking up information, there's no discernible difference, and no need to pay.

Furthermore, one must question why users would be willing to pay.

I noticed that Doubao focuses its charging on PPT creation, data review, and film and television creation, but how many of Doubao's user base actually utilize these features? It might be prudent to examine backend data before proceeding, lest there be a mismatch.

If we set aside emotions and break free from the expectation of everything being free, is Doubao's pricing excessive?

Looking at the global AI landscape, subscription models have long been the norm among industry giants. Comparing the top three players horizontally, one finds that the market's pricing logic is far more ruthless than imagined.

Claude's top-tier Pro version offers a 5x quota, with a stable monthly fee of about 900 yuan, valuable for developers.

ChatGPT Plus has differentiated into a higher-frequency Pro tier, with a 5x quota compared to Plus, starting at about 1,000 yuan per month.

Moreover, OpenAI was the first to abandon blind user growth and instead implement a paid model. It deeply bundled the o1-pro full-power model with the Pro plan, leaving free users with an ad-supported, limited version.

Gemini, Google AI Pro, starts at 145 yuan per month, while the Ultra version, deeply integrated with Google Workspace, costs nearly 2,800 yuan per month on an annual basis. While this may seem exorbitant, from the perspective of a cloud computational asset package, it's a viable choice for many small and medium-sized startups.

Doubao's model matrix capabilities are commendable, and under equivalent computational loads, offering a professional version at 500 yuan per month actually seems cost-effective.

After the stress test, introducing a tier priced at one or two hundred yuan would likely be acceptable to users.

As of March 2026, Doubao's daily token consumption has reached 120 trillion, a thousandfold increase from two years ago. Behind every line of generated code and every animated image lie tens of thousands of H100 GPUs consuming power.

Thus, Doubao has taken the lead in asserting its pricing power, serving as a mirror for the industry. Now, it's the domestic competing large models that should be torn:

If they follow suit and charge, users will immediately churn; if they maintain free access, they implicitly admit their model's value is low, condemning themselves to struggle in the shallow waters of traffic while failing to foster true productivity. Free, superficial chatting cannot drive logical evolution or long-text reconstruction in large models, only making them dumber and more entertainment-focused.

The fans in data centers continue to roar, burning real money every second.

This is not just ByteDance's bill but a choice left to every AI company in the AI era.

From: Benyuan Finance; Data support from Tianyancha