After hitting the $50 billion milestone, Kimi is facing two expectation management challenges.

A few days ago, Kimi released K3.1 Pro. 48 hours after the launch, Kimi issued an apology letter regarding its weekly usage quota.

On one hand, there is the carnival where its valuation skyrocketed from 4.3 billion USD to 500 billion USD in just 8 months. On the other hand, there are the bills of paying users. Two sets of expectation management targeting different groups collided at the exact same time window.

Kimi needs to answer or prove one question to the market: users are willing to keep paying for its fourth-ranked global benchmark score. After all, all valuations are ultimately determined by users' actual spending.

Two Sets of Expectation Management

"K3.1 is the top domestic model in my mind. But there are still two points to be improved: the speed is relatively slow, and the quota is still very limited."

"K3 gave people a stunning experience. K3.1 Pro has dense step arrangements, but there is still a chance that it will go down the wrong path. One wrong step will truly waste tokens and time."

These are some users' usage experiences after the release of K3.1 Pro.

Just like K3, the launch of K3.1 Pro is still a stunning presence among domestic large models. It scores 55 points on the Artificial Analysis Intelligence Index, ranking 4th globally and 1st among open-source models. Once the weight parameters are released, the gap between leading open-source and closed-source models will shrink from 12 points to 3 points, the smallest in history. But Kimi is facing pressure from two sets of expectation management.

The first one is valuation management.

From 4.3 billion USD at the beginning of the year to a surge of 500 billion USD in July this year, Kimi's valuation has multiplied more than 10 times in less than 8 months, making it a perfect model of capital driving premium for new technologies.

The fuel for this valuation rocket is very clear: by building large models that are continuously approaching the world-class level, it has elevated its technical position in the industry. This technical capability has almost become the anchor supporting Kimi's valuation.

For example, when its trump card K3 model was released in July, with 2.8 trillion parameters, the world's largest open-source model, it directly competed with Claude Fable 5 and GPT-5.6 Sol. With such a powerful weapon in hand, it had no worries about its future market territory, and the capital market's expectations were fully raised at that time. Now K3.1 Pro has been launched.

The second set of expectation management is reflected in user experience. This is an area that needs to be strengthened. After the release of K3.1 Pro, Kimi issued an apology letter.

The letter acknowledged two issues: rate limiting when members use K3.1 Pro, and "unnecessary quota consumption". The compensation plan is to refund 25% of the monthly quota, which is equivalent to the usage amount of one week.

A company valued at 500 billion USD publicly apologized because users "do not have enough quota". The two sets of expectations collided head-on at the same time window.

The Tension Behind the Two Sets of Expectation Management

The root of the tension between the two sets of expectation management may lie in the attributes of the product structure.

The core selling point of K3.1 Pro is its long-range Agent capability: independently breaking down tasks, invoking sub-Agents, and executing multi-step operations.

It scores 67% on Terminal-Bench 4.0, surpassing GPT-6 Astra's 58.2%.

However, the supporting logic behind such powerful technology is: the stronger the Agent, the faster the tokens are consumed. Every "intelligent decision" is consuming the user's subscription quota.

The data from Artificial Analysis is very harsh: K3.1 Pro outputs about 55 million tokens after running the full set of intelligence index tests, 14% more than K3, while the median of models in the same price range is only 87 million tokens. Converted, for the same task, the output volume of K3.1 Pro is about 2.5 times that of GPT-6 Astra.

Capital focuses on the ceiling, while users focus on the floor. International recognition is indeed rising. The title of the report by Ryan Fedasiuk, a researcher at the American Enterprise Institute, is "China Has Caught Up with Cutting-Edge AI", stating that the Sino-US gap has "collapsed from 6-8 months to a few weeks".

On Hacker News, the discussion post about K3.1 Pro received about 850 points and nearly 500 comments. The most popular discussion is that "Kimi is more straightforward and more aggressive than US models".

However, these popularity and good reputation belong to the developer community and evaluation ranking lists. The feelings of ordinary subscribing users are different from these, with a clear gap. As paying members, they care more about "why my quota is used up again this month".

Besides, in China, all major models have already launched price wars.

Zhipu's GLM-5.2 was released one month earlier, with a price of 1.4 USD per million input tokens and 4.4 USD per million output tokens, while K3 charges 3 USD per million input tokens and 15 USD per million output tokens.

Recently, DeepSeek also released V4.1 Flash-1017, with a completely opposite idea: instead of competing for the upper limit, it optimizes performance, speed and output length at the same time, while keeping the price unchanged.

These moves by competitors are undoubtedly accelerating the scramble for the user base.

In June 2026, Kimi had 7.29 million monthly active users, ranking 9th among all AI-native apps. This makes it even more necessary for Kimi to take good care of its more than 7 million users.

On a larger scale, the retention of paying users is crucial to the Pre-IPO story. Their willingness to pay is directly related to the valuation pillar of whether ARR (Annual Recurring Revenue) continues to grow.

It is worth noting that the external environment is also bringing stress tests to a large number of domestic AI companies.

Just two weeks ago, Anthropic's IPO valuation expectation was lowered from 2 trillion USD to 1.5 trillion USD, because its Q3 revenue quarter-on-quarter growth rate dropped from 146% to 45%.

The global capital market's patience with AI companies is shifting from "focusing on growth" to "focusing on delivery" (that is, focusing on solid operating results, not just the growth rate).

Zhu Xiaohu from GSR Ventures put it more bluntly: "The previous AI valuation was narrative-based, and it will shift to fundamental-based valuation in the future." All these serve as a reminder from the external market for Kimi.

The Balance Point Is Not About "Saving Quota", But About "Making It Clear"

Kimi took quick compensation actions, and the apology letter was quite sincere. However, users will not only look at this incident once, Kimi needs to prevent the same situation from happening again.

Kimi also stated that in the future it will "enable the same quota to complete more practical tasks". But what some users care more about is that they need to know exactly where their quota is spent and what they get in return. For example, how much is used for search, how much for analysis, which steps are necessary, and which steps can be cut off.

In fact, this challenge faced by Kimi is not unique to itself, but an industry-wide problem.

The professional paid version of Doubao has covered price ranges from 68 yuan to 500 yuan per month, but when users open their bills, they still only see a percentage, not "how much this in-depth research cost, which steps in the Agent task are necessary, and which steps are extra actions added by the model itself".

Zhipu teaches users to "reduce repeated confirmation" and "state all requirements clearly at one time" to save points. This direction is correct, but the responsibility of saving is still pushed to users, instead of the platform making the accounts clear for users.

Platforms like Coze clearly write down the unit price of tokens and the billing formula, but that is for developers. Ordinary subscribing users still face a billing link that they cannot understand and cannot interfere with.

In other words, if Agent capability is the barrier that distinguishes Kimi from Doubao and Qianwen, then Kimi's direction of managing user experience expectations should not be "you need to save usage later", but "make usage transparent and understandable".

Users are the real anchor behind the capital inflow. What a Pre-IPO company needs to prove to the capital market is not how high the model's benchmark score is, but users' willingness to keep paying for that score.

To borrow the popular nonsense line from the current hit drama Jewel in the Palace: "Relying only on evaluations from intermediary institutions is not enough. You need sincerity to make the soybean paste soup taste better and better."

This article is from the WeChat Official Account "Muhe Business and Finance", author: Gong Zheng, editor: Yang Jing, published with authorization from 36Kr.