After the Release of Kimi K3.1 Pro: A $50 Billion Valuation and an Apology Letter to Users Over Quotas

Recently, Kimi released K3.1 Pro. Forty-eight hours after the release, Kimi issued an apology letter over the weekly quota.

On one side is the frenzy of a valuation that surged from $4.3 billion to $50 billion in eight months. On the other is the bill of paying users. Two exercises in expectation management, aimed at different audiences, collided in the same time window.

Kimi needs to answer, or prove, one question to the market: that users are willing to keep paying for the world's fourth-highest benchmark score. After all, every valuation is ultimately decided by users' bills.

Two exercises in expectation management

"K3.1 is, in my mind, the best in China. But there are still two things to improve: it's fairly slow, and the quota is still very small."

"K3 felt stunning. K3.1 Pro takes very dense steps, but there's still a chance it goes down the wrong path, and one wrong step really wastes tokens and time."

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

Like K3, the release of K3.1 Pro remains a stunning presence among Chinese models. It scores 55 on the Artificial Analysis Intelligence Index, fourth globally and first among open-source models. Once the weights are released, the gap between the open-source and closed-source frontier will shrink from 12 points to 3, the smallest in history. But Kimi is facing pressure on two fronts of expectation management.

One is valuation management.

From $4.3 billion at the start of the year to a surge to $50 billion in July this year. Kimi's valuation has more than multiplied tenfold in less than eight months, a perfect template of capital driving a premium for new technology.

The fuel for the valuation rocket is clear. By using large models that keep approaching the international frontier, it raises its own technological position in the industry. This technical capability has almost become an anchor supporting Kimi's valuation.

For example, when the killer K3 model was released in July, with 2.8 trillion parameters, the world's largest open-source model, directly benchmarked against Claude Fable 5 and GPT-5.6 Sol, with the weapon in hand, the future territory was not a worry, and expectations in the capital market at the time were maxed out. And now K3.1 Pro has been released.

The second exercise in expectation management should be reflected in user experience. This is the point that needs strengthening. After the release of K3.1 Pro, Kimi sent an apology letter.

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

A company valued at $50 billion publicly apologizing because users "don't have enough quota." The two expectations collided head-on in the same time window.

The tension behind the two exercises in expectation management

The root of the tension between the two exercises in expectation management may lie in the structural attributes of the product.

K3.1 Pro's core selling point is long-horizon Agent capability: autonomously breaking down tasks, calling sub-Agents, and multi-step execution.

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

But behind such formidable technology, the supporting logic is: the stronger the Agent, the faster tokens burn. Every "intelligent decision" consumes the user's subscription quota.

Artificial Analysis's data is brutal: running the full Intelligence Index test suite, K3.1 Pro outputs about 55 million tokens, 14% more than K3, while the median for models at the same price point is only 87 million. Converted, for the same task, K3.1 Pro's output is about 2.5 times that of GPT-6 Astra.

Capital looks at the ceiling, users look at the floor. International acclaim is indeed warming. The title of a report by Ryan Fedasiuk, a researcher at the American Enterprise Institute, is exactly "China Has Caught Up to the AI Frontier," saying the US-China gap has "collapsed to a few weeks" from 6-8 months.

On Hacker News, the discussion thread on K3.1 Pro got about 850 points and nearly 500 comments, with the hottest discussion being that "Kimi is more blunt and more combative than American models."

But this buzz and reputation belong to the developer community and the benchmark leaderboards. Ordinary subscription users' feelings have a gap with these. As paying members, they care more about "how come my quota is gone again this month."

Moreover, domestically, the various models have long since started a price war.

Zhipu GLM-5.2 was released a month earlier, at $1.4 per million input tokens and $4.4 for output, while K3 is $3 for input and $15 for output.

Recently, DeepSeek also released V4.1 Flash-1017, with a completely opposite approach: not competing on the ceiling, but optimizing performance, speed, and output length together, at an unchanged price.

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

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

More broadly, the retention of paying users is crucial to the Pre-IPO story; their willingness to pay directly bears on whether ARR keeps growing, the pillar of the valuation.

Just two weeks ago, Anthropic's IPO valuation expectation was cut from $2 trillion to $1.5 trillion. The reason was that third-quarter revenue growth fell from 146% quarter-over-quarter to 45%.

GSR Ventures' Zhu Xiaohu put it more bluntly: "AI used to be narrative valuation; from now on it will shift to fundamental valuation." For Kimi, all of this can serve as a kind of outside reminder.

The balance point is not in "saving quota," but in "explaining clearly"

Kimi's compensation action was fast, and the apology letter was fairly sincere. But users won't look only at this one time; Kimi needs to prevent having to send another one next time.

Kimi also said that in the future it will "let the same quota complete more actual tasks." But what some users are more concerned about is needing to know further where the quota was spent and what it bought. For example, how much was used for search, how much for analysis, which step was necessary, and which step could be cut.

In fact, this challenge for Kimi is not unique to it, but an industry-wide problem.

Doubao's professional paid version has already been rolled out at price points of 68 to 500 yuan per month, but when users open their bill, what they still see is only a percentage, not "how much this deep research cost, which steps in the Agent task were necessary, and which steps were the model adding drama on its own."

Zhipu teaches users to "reduce back-and-forth confirmation" and "state requirements clearly in one go" to save credits. This direction is right, but the responsibility for saving is still pushed onto users, rather than the platform settling the bill clearly for users.

Platforms like Coze do write the token unit price and billing formula very clearly, but that is for developers; ordinary subscription users still face a billing chain they can neither understand nor intervene in.

In other words, if Agent capability is Kimi's moat distinguishing it from Doubao and Qwen, then the direction of Kimi's expectation management for user experience should not be "use it sparingly from now on," but "use it with clarity."

Users are the stabilizing anchor hidden behind the capital's increased bets. 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 a line of nonsense literature from the currently hottest "Dae Jang Geum": "Relying only on the evaluations of intermediary institutions is not enough; it takes sincerity to make the soybean paste soup better and better."

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