Don't take 345 million as your confidence.

The more I analyze the practice of burning tokens to attract users, the more I think it's like the top donor in a live - streaming room. When the donor is showering gifts, the screen is filled with "Big Brother, Big Brother." But once the tipping stops, the so - called affection fades to zero.

01

Doubao has a monthly active user base of 345 million, ranking first among domestic AI applications by a wide margin. People outside keep using this figure as evidence, and the core message is: Doubao has the confidence to start charging.

I don't see it that way. This 345 - million monthly active user base is not Doubao's confidence for charging; instead, it's the biggest obstacle to its charging.

Why? To clarify this, we first need to ask: Where do these 345 million users come from?

Actually, users mainly enter through three channels.

The first is the diversion from the Douyin ecosystem.

Doubao is embedded in more than 50 internal ByteDance business scenarios such as Douyin, Toutiao, and Fanqie Novel. Users never search for "I'm looking for an AI tool." They just click on it casually while swiping short - videos, not even knowing what they've opened.

The second is pre - installation on mobile phones. Through cooperation with terminals like OPPO and vivo, Doubao is directly on your desktop when you turn on the phone.

The third is social spread. In the early days, many people used Doubao for the first time because younger family members in the relative group taught their uncles and aunts to install it.

These three channels have one thing in common: from the day users enter, no one tells them "this thing is valuable."

Comparing with Kimi makes it clear:

Kimi is promoted through real - money advertising. It runs splash ads on Douyin and cooperates on evaluations on Bilibili. Every user reached has at least experienced the awareness that "AI has a price."

Users acquired through advertising, whether they end up buying or not, have already planted a thought in their minds: This thing costs money.

As for Doubao, this step doesn't exist at all.

I checked the user portrait of QuestMobile in Q1 2026. Nearly 70% of Doubao's users are in the 25 - 40 age group in the workplace. At first glance, it looks quite good.

Digging deeper, users in first - and second - tier cities only account for 45%. Instead, there are more users in the sinking market. Third - tier cities account for 24.6%, ranking first among all city levels.

I also noticed a detail in the Q1 survey. The penetration rate of Doubao's voice call function is approaching 40%, and it's particularly popular in fourth - and fifth - tier cities. More and more middle - aged and elderly users encounter AI for the first time in their lives through the zero - threshold way of "making a phone call."

So how much time do they spend on Doubao?

According to QuestMobile's data, the average daily usage time is less than 20 minutes. For comparison, voice - chat AI products can reach an average daily usage time of 70 - 110 minutes. What does less than 20 minutes mean? Open it, ask a question, get the answer, close it, and leave.

This is the foundation of the 345 - million monthly active user base. They are attracted by free offers, enter through the lowest - threshold way, stay for a short time, and have never established a payment expectation from the beginning.

If you try to charge this group of users, it's like a park that has never charged an entrance fee suddenly putting up a sign at the gate: "The scenic area has been upgraded. Please buy tickets to enter."

The problem is not whether the ticket is worth it. It's that the people who enter don't think this place needs a ticket from the first day.

02

Doubao Pro started charging on June 24th, but as early as May 4th, the payment plan had already been updated in the app store.

It's like spreading the news two months in advance.

I checked the statistics of the AI data platform aicpb. In May, Doubao's monthly active users dropped by 6.1 million, from 336 million to 330 million, a month - on - month decrease of 1.81%. This is the first time such a large - scale user exodus has occurred since Doubao was launched in 2023.

Of course, a one - month fluctuation doesn't explain everything. But when we look at this phenomenon in the context of the entire industry, it means something different.

6.1 million in a 300 - million user base is indeed a small proportion. What's really worth pondering is the timing of these people leaving.

They haven't used the paid version, haven't compared the differences between the free and paid versions, and haven't even seen what the payment entrance looks like. Just the five words "heard it will charge" are enough for them to uninstall.

You must be wondering, is it really necessary?

It is. Because you've never mentioned money to them from the beginning. You've invited them for free for three years, and suddenly you say you'll start charging. Their first reaction is just four words: "Why should you?"

They haven't even reached the stage of evaluating the product's value, and they've already rejected you emotionally.

What happened during the same period? The monthly active users of Alibaba Qianwen soared by more than 13 million. Qianwen and Baidu Wenxin clearly stated that they will continue to be free, and Yuanbao also publicly stated that it has no charging plan.

Just as you've shown the price tag, the one next door has removed the door and welcomes you.

In a market where the switching cost is almost zero, users' choices are very simple: if you charge and the one next door doesn't, then I'll go to the next door first and take a look.

Then I ask you, after the professional version is really launched, what good things can the paid version offer to bring back the hesitant users?

I've looked through a lot of information and found that most people mention these things:

The paid version only has a little more functionality than the free version, but paying doesn't necessarily mean better results. The results depend on the model itself, not on whether you pay or not. The core value, to put it simply, is a few more usage quotas and priority queuing during peak hours.

This isn't what I'm saying; it's what the customer service said.

What's more interesting is that when the professional version was launched, the free version was also upgraded to Doubao 2.1 Turbo. Since the free version is getting stronger on its own, users can't see a reason to spend money.

Think about it. If the free version is upgraded every time, then the "more quotas" of the paid version will always be just a quantitative difference, and there will never be a qualitative leap.

I've used it for a while. Well, the free version can answer basic questions, but it gets stuck on slightly more complex tasks. For example, information verification. I often use data when writing articles. When I throw dozens of data points at it, it just gets confused.

What's the most frustrating feeling?

The paid version doesn't give me a reason to say "I can't do without it." The tasks it can do, a free AI can do the same.

I also saw a summary of user feedback from the Doubao Pro internal testing group. As of early June, paid users' complaints mainly focused on three things: lack of personalized functions, little difference from the free version, and poor overall experience.

One user put it very accurately. The professional version should offer something truly irreplaceable, such as exclusive access to the flagship model, dedicated server channels, and a larger context window. These are worth the price.

To put it simply, what the paid version of Doubao offers now is "a little more quantity," more usage quotas and more queuing opportunities. What users want is "a qualitative difference," that is, the paid version can handle things that the free version can't.

This gap between quantity and quality is the real reason why those 6.1 million people turned away at the "hearing" stage.

Their intuition told them that this product probably couldn't give them a reason to spend money. Later, when the professional version was really launched, their intuition was right.

03

By now, you must be scolding me: You're just a hater who keeps criticizing Doubao, right?

Don't be in a hurry. Chinese users are willing to pay for AI. I pay for it every month myself, and I'm also a paying user of several domestic products. The problem is, the way of paying is different.

Let me tell you about a product that has raised its price three times this year. Instead of leaving, users are rushing to buy it.

Zhipu's GLM Coding Plan has a single function: AI - assisted code writing.

I checked its price changes. In February 2026, Zhipu issued a price adjustment letter, canceling the first - purchase discount, with an overall price increase of at least 30%.

The reason was "the continuous strong growth of market demand, and the rapid increase in user scale and call volume." After the adjustment, the domestic pricing became Lite at 49 yuan, Pro at 149 yuan, and Max at 469 yuan.

This was the first price increase.

In April, when the GLM - 5.1 model was released, the API price was raised by another 10%. Immediately afterwards, the monthly payment price of the overseas version of the Coding Plan almost doubled, which was the third price increase.

After three price increases, Zhipu's CEO Zhang Peng released a set of data during the earnings conference call: In Q1 2026, the interface call price increased by 83%, but the call volume increased by 400%.

The price increased by 83%, and the call volume increased by 400%. Think about it. This is not normal supply - and - demand; it's like the product is out of stock.

This is indeed the case. The Max tier of Zhipu's Coding Plan is often sold out. Some packages have been changed to a limited - quantity purchase at 10:00 every day. If you're slow, you won't get it.

Moreover, it's not just Zhipu. The keyword for the entire AI programming subscription track in 2026 is price increase.

MiniMax, Kimi, Alibaba Bailian, Baidu Smart Cloud, and Huoshan Ark have all joined in. At least seven platforms are offering Coding Plans. Since the second half of 2025, the industry's unified actions have been to tighten free quotas, cancel first - purchase discounts, and switch to more detailed Token billing.

MiniMax even publicly wrote two apology letters this year because of the sudden change in the billing method. Interestingly, the apology was because the price increase changed the rules and users had objections, but they still kept using the product.

It's very interesting to compare this with what's happening to Doubao.

Doubao hasn't even officially started charging, but 6.1 million users left as soon as they heard the news. Zhipu has raised its price three times, and users are lining up to buy. In the same market, among the same group of Chinese users, and during the same period, why is there such a big difference in their reactions?

The problem is not the product's ability. Zhipu's model is not necessarily stronger than Doubao's. In many benchmarks, Doubao 2.1 Pro is on the same level as GLM - 5.1.

The difference lies in earlier things.

From the very beginning, Zhipu's Coding Plan targeted developers. When these people came in, they had a very specific need:

I just want to write code. I need a tool that is cheaper than Cursor but sufficient for my needs. They knew from the start that this thing costs money. They are evaluating "which one offers better cost - effectiveness."

What about Doubao's users?

Think about how these users entered. At the moment they entered, the word "payment" never crossed their minds.

When it comes to charging, one group scolds but still buys, while the other group wants to leave before the charging even starts. The difference lies in what's in users' minds when they enter the door.

04

By now, you might be thinking: Since the problem lies in how to attract users, can't Doubao change the way to educate users?

Of course it can, and it's already doing it.

Now when you go to the airport, you can see Doubao's new billboards everywhere, saying "Make Doubao your new work habit, let it do the work first."

But think about it carefully. This billboard itself is a symptom. When have you ever seen WeChat advertising at the airport saying "WeChat can also be used for office work"? It's not necessary. Users know it themselves.

When have you ever seen DeepSeek spending money to shout everywhere "I can write code"? It's also not necessary. Developers will vote with their feet.

Doubao spends money on billboard space to tell others "I can really do the work," which only shows one thing: No one thinks so about it now.

Not only do I think so. If you look at the AI product comparisons that come out on major platforms from time to time, these comparisons gradually form a mental map:

Choose DeepSeek for hardcore technology, Kimi and Baidu Wenku for office work, Doubao for life and entertainment, and Zhipu and MiniMax for AI programming. No one organized it, and no one discussed it. It's all chosen by users themselves.

Doubao has already been placed in the "life and entertainment" category in the minds of 300 million people. You're not unable to do PPTs, data analysis, or code generation. All the functions are there. The problem is that the label has been stuck for a long time, and users don't believe it.

Coincidentally, the main selling point of this charging is the "office task mode," including operating the local computer, invoking the browser, and having a built - in Office suite. Think about it. The brand image is "entertainment," and the charging scenario is "office." Isn't this a contradiction?

In users' minds, Doubao is the thing that randomly replies to messages, teaches grandmas to chat, and creates P - star selfies for the WeChat Moments. Now you suddenly ask it to dress up in a suit and do work for you. Their first reaction is: Is it reliable?

Let me make a comparison for you.

When ChatGPT Plus was launched in February 2023, OpenAI did something: Only paid users could use GPT - 4, while free users could only use GPT - 3.5.

Note that the free version and the paid version are not the same model at all. People who have used the free version can clearly see the gap when they see screenshots of GPT - 4 posted by others.

This is one way: Lock the best things behind the paywall from the start. The six - month payment retention rate is 71%, ranking first among global AI subscription products.

The other way is to offer everything for free from the beginning. After three years of free use, suddenly say you'll start charging. Before the official charging even starts, the monthly active users have already dropped by 6.1 million.

Two ways, two outcomes. The difference lies in the most fundamental thought in users' minds: Is this thing really worth the money?

I checked the data from Earnest Analytics. The six - month payment retention rate of ChatGPT Plus is 71%. Where does this retention rate come from? From the moment users enter, they are taught one thing:

The free version and the paid version are two different things. If you want to use my new model, you need to pay. This awareness was planted early on, while for Doubao, the early - planted awareness was "free."

Let's look at two other products within ByteDance. Jimeng is for AI video generation. Users know from the start that it's a creative tool, and it's obvious how much computing power is needed to generate a video. Charging by points is perfectly reasonable.