Beanbao will start paying in late June-with 345 million monthly users, the era of free AI in China is really about to end

Let me put out the original text of today's news first--

On June 1, 36Krypton exclusively disclosed that bean buns are expected to officially launch paid content in late June, and relevant functions will be updated at the Force Power Conference held during the same period.

This news appears bland. But if you are like me and list all the changes in the domestic AI price ecosystem in the past four months by time, you will find that this is actually a matter of quite structural significance.

It's not an isolated "company that decides to collect money." It is the last piece of the puzzle in a series of price reversals.

Together, this series of actions tells us one thing-the era of "AI free for all" that China users have become accustomed to in the past three years is completely coming to an end.

Next, I will put out all the data one by one, and then tell the judgment I see.

Let's first know what kind of product bean buns are

This part of the data is the important root cause of this incident. Collecting money for an AI product that lasts for 1 million months is not a big deal. But bean buns are not.

According to QuestMobile's data for the first quarter of 2026, there are 345 million monthly active users of bean bags, making it the first AI native application in China. During the same period, 166 million thousand questions and 127 million DeepSeek. Doubao family is almost equal to the sum of the second and third places.

The global data released by the AI product list in April is even more dazzling-the monthly activity of bean bag apps is about 336 million, ranking second in the world, second only to ChatGPT. This is the first time that China's AI applications have reached this level of magnitude in the global list.

During the Spring Festival week of 2026, Doubao relied on its status as the exclusive AI cloud partner of the Spring Festival Gala, and its daily activity peak reached 145 million, becoming the first universal AI application in China to enter the "100 million-level daily activity".

The overseas version is called Dola (formerly known as Cici). In the first quarter of 2026, there were 72 million downloads in a single quarter, exceeding 200 million times. The average daily activity in March was close to 12 million.

The back-end data is even fiercer-Volcano Engine itself disclosed that the average daily Token usage of the Doubao model has exceeded 50 trillion yuan, an increase of more than 10 times over the same period last year.

Deciding to charge for products of this magnitude will drive the entire industry to follow suit.

What exactly should I charge for bean buns and how should I collect them?

There are many details in this exclusive report on 36 Krypton, so I will go through the key items.

On the timetable-it will take about one month for the bean bag PC and mobile terminals to complete the adaptation and transformation of basic functions and charging systems, so it was chosen to go online in late June. If everything goes well, the third quarter will further combine e-commerce function updates to drain Douyin Mall through subsidies. The fourth quarter entered the operating period. The real pace of the whole thing is to prepare for commercial returns in 2027 and beyond. Therefore, in 2026, Doubao will not use the penetration rate of paying users as an evaluation indicator.

Paying function positioning-In a report by China Financial News, people familiar with the matter mentioned that the payment function of bean buns mainly focuses on complex tasks and productivity scenarios, such as PPT generation, data analysis, and film and television production. The common characteristics of these scenarios are that they consume a lot of computing power and long reasoning time. Simply put-it's okay for you to use basic chat for free, but if you want Doubao to help you make a 30-page PPT, you have to pay.

In fact, the paid subscription plan for bean buns was low-key updated in the App Store on May 4-the standard version is 68 yuan per month, the enhanced version is 200 yuan per month, and the professional version is 500 yuan per month. There was a wave of discussion at that time. The Doubao official immediately responded that "Doubao always provides free services, exploring more value-added content on a free basis."

So strictly speaking, the matter of paying for bean buns was foreshadowed in early May, and the tone was only officially set in late June.

!

But if you look at it alone, it will not be so meaningful

To see clearly the true weight of the payment for bean buns in late June, we must take a look at all the actions of the domestic AI price ecosystem in the past four months.

I suggest you read the timeline below from beginning to end. After reading it, you will have the feeling of "I'm going to this matter actually started a long time ago."

February 12 --Intelligence released GLM-5. At the same time, it announced the structural adjustment of the GLM Coding Plan package and the cancellation of the first purchase discount, with the overall increase starting from 30%. This is the first time a domestic model company has made a clear price increase this year.

March 11 --Tencent Cloud agent development platform increased the price of the Mixed Origin Series Model API. Tencent HY2.0 Instruction was adjusted from 0.0008 yuan per thousand Tokens to 0.004505 yuan per thousand Tokens, an increase of more than 460%. This knife is quite heavy.

March 16 --Intelligent Vision released GLM-5-Turbo, simultaneously raising the API price by 20%. Compared with 4.7, the cumulative increase was 83%.

April 8 --Intelligence released GLM-5.1. API prices will be increased by another 10%. This is the third time that Zhipu has increased its price during the year. On April 12, the monthly payment for the overseas version of Smart Coding Plan almost doubled. In the dimension of cache hit prices, GLM-5.1 has already equalled Anthropic Claude Opus 4.5-"Domestic model prices are aligned with overseas prices for the first time."

April 18 -Alibaba Cloud and Baidu Intelligent Cloud announced on the same day that they would increase the prices of AI computing power, storage and other products, an increase of 5% to 34%. Alibaba Cloud also canceled the basic package of Bailian Platform and no longer provided the minimum version of services. It is no coincidence that three China cloud manufacturers raised prices on the same day.

May 9 --Tencent Cloud AI computing power and other products will be increased by another 5%.

May 31 --DeepSeek V4-Pro API will be permanently reduced to 1/4 of the original price. The new price is 0.025 yuan per million Tokens for input cache hits, 3 yuan for input misses, and 6 yuan for output. This is the only domestic manufacturer that has clearly followed the route of "reverse price reduction".

June 3 --GitHub Copilot globally switched to a new pay-by-Token model and introduced the "GitHub AI Credits" virtual currency. The overseas programmer community exclaimed,"A few hours of AI use will consume a large chunk of the monthly quota."

Late June --Official payment for bean bags + Force conference.

You arrange these nine items according to time--

February to June, 4 months. The domestic AI price ecosystem has completed a complete reversal. The reversal from "robbing users for free regardless of cost" to "screening customers based on value". The brake on the monthly income of 345 million yuan in bean buns is the loudest sound of this wave of reversal.

!

Why did this incident happen at this time?

After reading the timeline above, my first reaction was "Why are we all crowded in these months?"

To straighten it out, the root cause is two things.

First thing, the cost of computing power has really increased.

The October 2025 low for the NVIDIA H100 one-year GPU lease contract is $1.70 per hour per GPU. By March 2026, it has risen to $2.35 per hour per GPU. The increase is close to 40%.

At the end of March, Huawei's Shengteng server series launched the second round of price adjustments.

The income of large domestic model companies comes from "selling APIs" and the cost comes from "purchasing power." If the computing power side increases by 40%, the selling price side does not move, the gross profit will be eaten up. This is the most direct physical basis for "cloud vendors + model companies" to increase prices together.

In DeepSeek's announcement on May 31, it added one sentence-"The Pro version is limited by high-end computing power, and the current service throughput is relatively limited. It is expected that after the Shengteng 950 supernode is launched in batches in the second half of the year, the price of the Pro will be reduced significantly."

It tells you bluntly that the prerequisite for lowering prices is to have cheap computing power. Without cheap computing power, no one can reduce it.

Second, the Agent verifies that the user is willing to pay.

OpenRouter platform data-Token consumption of Agent products such as OpenClaw soared from 80.6 billion on February 3, 2026 to 358 billion on March 4 within one month, an increase of approximately 4.4 times.

Agents often consume millions of Tokens for a single task, far exceeding the usage of traditional chat. Tools like Claude Code have been verified overseas-users are willing to pay for an Agent who "can help me complete a complete task." Hundreds of thousands of users around the world are already paying for a $200 monthly Claude Code subscription.

This incident directly rewritten the logic of the model company. Price increases are no longer "to pass on costs to users." The core purpose of the price increase is to "screen out low-ARPU users who are just here to collect free wool, and leave behind heavy Agent users who are willing to pay."

What Zhipu CEO Zhang Peng said at the 2026 Zhongguancun Forum Annual Meeting-"Adjusting prices is to return to normal commercial value, and long-term low-price competition is not conducive to the development of the industry. The bottleneck lies in computing power, not in customers."

This is the official farewell to the three-year logic of "free market exchange".

The true weight of the ## bean bag payment is here

Now back to the news on June 1st, 36 Krypton.

The monthly income of bean buns is 345 million yuan, and the average daily Token is 50 trillion yuan. Based on the industry's average "reasoning cost per thousand Tokens of 0.001 to 0.005 yuan", the daily computing power expenditure of bean buns is an expenditure of tens of millions of yuan. This part of the money has been burning for almost two years, and it will be obviously under pressure in 2026.

There is a detail in the original Krypton text that is the most noteworthy in this news--

"Since April 2026, users 'acceptance of AI shopping recommendations has increased compared with the second half of last year. The conversion rate of clicking on the product card and viewing the details page has reached more than 3%. On the other hand, the current user growth of Doubao has slowed down. According to 36Krypton, part of the slowdown in growth is also the result of Byte's intentional actions-expensive computing power and reasoning costs have caused Byte to reduce its investment in promoting bean buns in the new year."

You put these two sentences together--

Byte himself is actively stepping on the brakes. Change the strategy of "grabbing users regardless of cost" in the past two years to "reducing promotion + increasing payment conversion + connecting to Douyin e-commerce."

This is a systematic shift within Byte to the beanbao business model.

What is the weight of this matter--

It is not a byte-family judgment. It is the last piece of the jigsaw puzzle in the domestic AI industry's collective reversal in the past four months. Once all products of the magnitude of bean buns are announced to pay, the remaining small and medium-sized manufacturers will basically only have to follow suit.

This means to you-the good days of "opening an AI tool and using N advanced functions for free" in the past three years will basically end in the second half of 2026.

So what exactly do you want to do

If you use AI products such as Doubao, Qianwen, DeepSeek, Kimi, Wenxin, and Zhipu Qingyan every day, four specific things will happen in the next 6 months. I'll tell you one by one.

First, the free layer will become thinner.

The official statement of Doubao is that "free services continue, and paying is value-added." There is a high probability that other manufacturers will follow next. But the actual boundary of the word "value-added" will become wider and wider-scenarios such as PPT generation, long document analysis, complex code, video generation, and multi-step Agent tasks will all be added to the payment layer bit by bit. What the free layer will leave behind is "cheap" functions such as simple chat, basic translation, and short text summary.

If you used to rely on free AI to solve heavy tasks at work, such as "ask AI to help me write this PDF into a PPT" and "ask AI to help me translate and subtitle this video", there is a high probability that you will have to start paying in 6 months.

The second thing, billing by Token is coming.

The three files (68, 200, and 500 yuan per month) already available on the Bean Bag App Store are subscription based. But you see that GitHub Copilot just switched to a new pay-by-Token model on June 3-how expensive it will be in the future, and how cheaper it will be to use less. There is a high probability that domestic manufacturers will follow in the next 6 to 12 months.

The experience of billing by Token will be very different. You used to subscribe to 68 yuan per month and you could use it without brain. In the future, you might run an Agent task and eat up 30% of your quota. Mental accounting is completely different.

Third, DeepSeek will become a safe haven for "free parties".

In the short term, DeepSeek is the only company in the domestic mainstream model that clearly follows the "extremely cheap + open source base" route. Liang Wenfeng publicly promised to continue to do open source. The company's main goal is to "promote the technological boundary, not realize it as soon as possible."

If you're not planning on paying, DeepSeek will be the most stable option for the next 12 months. The premise is that you can accept its limitations-the Pro version currently has limited service throughput due to limited high-end computing power. The response speed is slower than Doubao and Qianwen. In the second half of the year, the Shengteng 950 supernode will drop again after it is launched in batches.

Fourth, the entrepreneur's accounting logic needs to be redone.

If you are doing an AI application entrepreneurship, the cost budget table for the past two years will have to be rewritten.

API prices have increased by an average of 30% to 80%(Smart Spectrum, Alibaba, Tencent, and Baidu), and computing power has increased by another 5% to 34%, which means that unit economics has been compressed across the board. Free users cannot afford it for a long time.

The entrepreneurial logic of AI applications in 2026 will have a high probability of changing from "first robbing users for free and then thinking about a way to monetize them" to "each user's LTV and CAC must be calculated clearly on the first day of launch." The road of burning money to rob users has gone through a wave in 2024-2025, and the road in 2026 has basically closed to small and medium-sized teams.

Last

On June 1st, the news of payment for the bean bag on 36 krypton was a product news item.

Digging down one layer is the last piece of the puzzle that the domestic AI price ecosystem has reversed in the past four months-Smart Spectrum in February, Tencent Cloud in March, Arijia Baidu in April, Tencent's computing power rose again, on May 31, DeepSeek reversed bargaining, on June 3, GitHub Copilot charged by Token, and in late June, payment by bean bag officially came into effect.

Digging further down is the gross profit space that has been compressed in the past 18 months between the revenue from "selling APIs" and the cost of "buying power" of large model companies. The one-year NVIDIA H100 lease has increased from $1.70 per hour to $2.35 per hour. Huawei Shengteng series has the second round of price adjustments at the end of March. Model companies cannot bear the burden of a 40% increase in computing power on their own forever.

Digging further down, after the Agent product verified that "users are willing to pay for the Agent", the entire industry began to screen customers with price-leaving those who are willing to pay and inviting those who collect wool for free.

This is the real way the era of "free AI for all" over the past three years has ended.

It is not a unilateral decision by a company. It comes from a structural turning point driven by three things: "rising computing power + Agent verification of paying propensity + no longer burning unlimited capital."

Stepping on the brake on the plate of bean buns, which earns 345 million yuan a month, is the loudest sound of this turning point.

In the next six months, your relationship with AI tools will slowly change from "It's fun to use for free" to "Is this feature enough to unlock paid files?" This is a big experience change.

Whether it is worth paying is up to you to decide for yourself.

But the time for reckoning has come.

!