Kimi K3.1 Pro Apologizes 48 Hours After Launch, Compensates Subscribers with 25% Usage

In just 48 hours, Moonshot AI's flagship model went from a high-profile launch to a public apology.

In the early hours of October 26, Kimi officially apologized for the rate limiting and abnormal quota consumption encountered by subscription members after K3.1 Pro went live, and offered a 25% usage compensation. Behind this "launch—apology" cycle lies the fundamental pricing challenge for large models in the Agent era—the cheaper per-token pricing gets, the more expensive users' overall usage costs become.

The compensation plan is as follows: all subscription members within their validity period will receive a usage compensation equal to 25% of their monthly total; if the currently used quota is less than 25% of the monthly total, usage will be directly reset to 0, and the quota progress bar will be adjusted within the same day. The company also admitted that "the problem is on us," and simultaneously lowered the per-token API pricing—it is precisely this combination of "cutting prices while rate-limiting and compensating" that makes the perceived gap for consumer users even more glaring.

User feedback concretely illustrates this anxiety: one subscriber on the 199 yuan tier said that merely having the model install a single Skill consumed 70% of the 5-hour limit, and that one hour of normal use used up 27% of the 5-hour rolling quota. The company simultaneously lowered per-token API pricing, yet for consumer users it became "more expensive," and quota anxiety has become the biggest reputational risk for this new model.

The deeper test lies in the fact that the full weights of K3.1 Pro will not be open until November 4. Before then, rate-limiting measures will continue, and if the compensation plan cannot stabilize user retention and brand trust within this window, every model upgrade may repeat the paradox of "the stronger the capability, the more complaints."

Installing one Skill burns 70% of the 5-hour limit

The high consumption of Agent tasks is supported by specific cases. A user subscribed to the 199 yuan tier said that merely having the model complete the simple task of installing a single Skill consumed 70% of the 5-hour rolling limit; another user calculated that one hour of normal use used up 27% of the 5-hour rolling quota, equivalent to 1.07% of the monthly total quota.

A more extreme case came from an Agent Swarm task: after the user launched a task to generate 100 gallery works, the five-hour window quota was exhausted in about one minute, the task produced no results, and the terminal continuously popped up HTTP 429 errors—the company noted that this was an isolated case involving a single user.

The China Academy of Information and Communications Technology's "Token Economy Development Research Report (2026)" offers an industry-level explanation: the token consumption of a single agent task can reach 10 to 100 times that of ordinary conversational applications, and a single 1.5-minute AI short drama episode consumes 5 million to 10 million tokens; domestic daily average token calls grew from 100 billion in early 2024 to 500 trillion by June 2026, an increase of more than a thousandfold. User demand has shifted from simple Q&A to engineering-grade complex tasks, yet the pricing system has not kept up.

The "scissors gap" between falling per-token prices and rising overall costs

The officially announced API pricing for K3.1 Pro is only one-fifth that of comparable flagship products, and the company simultaneously lowered the per-token API pricing. But in terms of the actual experience of consumer users, overall usage costs have risen rather than fallen.

Zhu Keli, founding president of the Guoyan New Economy Research Institute and chief expert on the intelligent economy, defines this contrast as a "structural industry change" brought about by the deployment of Agent intelligent agents, and judges that it "will be a phased norm for some time to come."

The logic is this: in the past, ordinary conversational interactions were brief and consumption was controllable, so the token unit price was basically equivalent to the usage cost; in the Agent era of long chains of thought and multi-tool calls, models need to autonomously break down tasks and repeatedly verify and iterate, so the consumption of a single complete task surges exponentially, and even if the unit price is lowered, the extra consumption brought by redundant reasoning still raises overall costs. Vendors continue to cut prices at the API layer to compete for the developer ecosystem, while consumer subscription users are paying for the high consumption of Agent tasks.

The monthly subscription model is reaching its end, and tiered pricing is emerging

Zhu Keli's judgment on the direction of the business model is quite direct: the traditional unlimited consumer monthly subscription model is unsustainable, the industry is bidding farewell to the era of extensive low pricing, and the mainstream in the future will shift to a tiered system of "basic monthly subscription as a floor + high-tier tasks billed by usage"—light daily interactions retain an affordable subscription, while complex Agent workflows are priced separately.

This aligns closely with Kimi's predicament. After K3 was released in July this year, user requests exceeded expectations in a short time, and the platform temporarily suspended new user subscriptions; the rate limiting and abnormal quota consumption of K3.1 Pro this time are essentially an upgraded version of the same problem. Zhang Youyu, a distinguished researcher at Peking University, believes the key to breaking the deadlock lies in Model Routing—using lightweight models to handle 80% of routine requests, directing high-end computing power to complex long-chain Agent tasks, and relying on scheduling to reduce overall inference costs.

English text is being prepared.

The trust window before weights are opened

The full weights of K3.1 Pro will be opened before November 4, and open source is seen as an important means of diverting computing pressure and expanding the developer ecosystem: after the weights are opened, a large number of enterprise-level inference needs can be deployed on their own clusters, easing the load on the original vendor. But before the window arrives, rate-limiting measures will continue, which constitutes a real test of user retention and brand trust.

This is not the first time Moonshot AI has faced supply-demand fluctuations; in July this year it temporarily suspended new user subscriptions because requests exceeded expectations. Overseas leading vendors OpenAI and Anthropic both have stable cloud computing foundations, while most domestic independent AI vendors lack their own cloud ecosystems; combined with constraints on high-end chip supply, the total available computing power is relatively fixed, and can only be dynamically allocated between model training and online inference. The stronger the model capability and the more heavy tasks it attracts, the greater the pressure on the inference side, and technological breakthroughs instead bring new resource tests.

Going forward, attention should be paid to the pace of adjustment of rate-limiting measures and actual changes in user retention around the opening of full weights on November 4. Risk disclosure and disclaimer

Markets carry risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this is at the user's own risk.