Doubao Finally Begins Charging Users

On May 4, ByteDance’s AI assistant quietly rolled out a paid subscription notice in the Apple App Store, introducing three pricing tiers at once: 68 yuan per month for the standard version, 200 yuan for the enhanced version, and 500 yuan for the professional version. The most expensive annual subscription totals 5,088 yuan.

The topic “Doubao Paid Subscription” quickly shot to the top of Weibo’s trending list.

The official response came swiftly: Basic features will remain permanently free, while the paid plans are still in the testing phase, mainly targeting high-computing-power productivity scenarios such as PPT generation, data analysis, and film and television production.

This messaging closely mirrors the standard pre-charge strategy of all internet products: “Keep using it; the free stuff isn’t going anywhere.”

But behind this move lies not just strategy, but a long-delayed yet ultimately unavoidable question: Can the AI business sustain itself?

Doubao boasts 345 million monthly active users, with daily active users once peaking at 145 million. It firmly dominates the consumer-end market, with a user base larger than the combined total of its second- to fifth-place competitors.

In the classical internet era, a product of this scale would have already become a major revenue generator.

But large AI models are not like Today’s Headlines.

For Today’s Headlines, each additional user comes with nearly zero marginal costs—just a few extra lines of code on the servers. For Doubao, every additional conversation consumes real electricity through GPU clusters. Rows of H800s and A100s eagerly await their next computation.

ByteDance’s capital expenditures are projected to hit 160 billion yuan in 2025, with roughly 90 billion yuan earmarked for AI computing power procurement. Doubao’s average daily token usage has soared past 120 trillion, a thousandfold increase since its launch.

Meanwhile, media reports indicate that ByteDance’s net profit in 2025 has plummeted by over 70% year-on-year due to sustained AI computing power investments. Another leaked research briefing suggests no clear path to commercialization yet, with inference costs pressuring profits.

A straightforward calculation leads to a brutal conclusion: If users keep consuming resources without paying, the company will collapse.

So Doubao had to act—but with careful planning.

By listing 68 yuan, 200 yuan, and 500 yuan tiers in the app store and adding the disclaimer that “plan details are still being tested,” this serves as both a stress test and a market probe. By introducing the highest price first, then offering a discount at launch, users feel like they’re getting a bargain. With the 500-yuan anchor set, other tiers seem more reasonable. This is a classic business tactic.

Nomura Securities believes Doubao, as China’s first AI chatbot to introduce paid subscriptions, marks a milestone for the market. Morgan Stanley notes the pricing “targets professional users,” aiming at creators and knowledge workers rather than the general public. However, Morgan Stanley also calculates that based on varying penetration rates, Doubao’s annual subscription revenue could range from 100 million to 1.5 billion USD. Even taking the midpoint, this remains modest compared to ByteDance’s core advertising revenue.

Paid subscriptions won’t become a major revenue stream anytime soon. So what is ByteDance really after?

Two motives: one overt, one covert.

The overt motive is “user screening.” Among over 300 million monthly active users, many are “computing power killers”—those running deep data analyses, generating ultra-long PPTs, or calling on film and television models. They consume far more electricity and computing power than average users. The paywall helps identify these users and make them bear their own costs while avoiding scaring off casual users who just want to chat or look up information.

The covert motive is more intriguing.

Doubao faces an awkward predicament: Most users have no idea how powerful it truly is. To control costs, the free version restricts context windows and reasoning depth, leaving ordinary users with the impression that Doubao is “convenient but not that strong.”

Yet ByteDance’s model matrix released over the past two years is no joke. Seed 2.0 Pro rivals GPT 5.2; Seedance 2.0 rivals Sora; and the programming model Seed Code. Doubao’s capabilities far exceed the average user’s perception of it as merely a writing and chatting tool.

The paid professional version aims to showcase these hidden strengths. Let willing payers see Doubao’s true value, while making unwilling payers understand that you get what you pay for.

The most striking twist may lie here.

In May 2024, Doubao slashed domestic API prices to “bone-crushing” levels, personally igniting a fierce price war. Two years later, after becoming China’s most widely used “free cafeteria” for AI users, it is now the first to stand up and close the free tap.

All stories of burning cash end the same way.

In the 1990s, Microsoft bundled Internet Explorer for free, crushing Netscape. In the early 2000s, China’s internet trifecta—portals, email, and search engines—used “free” as bait to redirect traffic to ads or premium services for monetization. During the mobile internet era, antivirus software went free, food delivery subsidies rained down, and ride-hailing coupons burned through cash.

Each round followed the same pattern: Capture market share first, then monetize; burn cash first, then clean up. In the end, someone always has to foot the bill.

The difference is that past expenditures were on servers and network cables—cheap after they burned out. Now, it’s on graphics cards and electricity, where each additional user means another hard cash outflow.

Tan Dai, head of Volcano Engine, once said: “The price difference in tokens essentially reflects the difference in their underlying capabilities.” This sounds elegant, but the reverse is equally true: If capabilities can’t justify prices, payers will walk away.

Currently, no large AI model has established an absolute moat in basic Q&A, leaving user loyalty as fragile as a windowpane. As soon as news of charging breaks, users will immediately seek alternatives.

The real chain reaction is that Doubao’s lead will trigger a domino effect of followers. Once tiered pricing takes hold, the entire domestic AI consumer market will shift from “whoever is free wins” to “whoever is worth the price wins.” Nomura Securities already predicts the entire market will eventually transition to subscription models.

From a harsher perspective: After burning cash in the model wars, even ByteDance is struggling to hold on.

This company, which once burned money to dominate short video, news, gaming, and other battlefields, now shows fatigue for the first time in the large model evolution cycle. Its computing power reserves, user scale, and capital thickness are unmatched domestically. Yet even so, it must end the “free cafeteria” it created.

Doubao’s paid subscription can be seen as a watershed: It marks China’s large AI model industry shifting from “who has the larger plate” to “who has stronger profitability.” The three-year-long cash-burning land grab in China’s large AI model sector has reached a temporary end. A bloodier final round has officially begun.

The new rules are no longer “who has more users” but “who has lower costs, better experiences, and a viable business model.” Small and mid-sized players lacking ByteDance’s computing power reserves, DeepSeek’s architectural breakthroughs, or OpenAI’s brand premium will be quickly cleared out.

The cruelty of the final round lies in its refusal to grant anyone preparation time.