Moonshot AI apologises for Kimi K3.1 Pro limits and hands back 25% of monthly quota

An apology and a quota reset

On October 26, the official Kimi account on Weibo, a Chinese social network, published a letter to users. It apologised for rate limiting and for quota being used up unnecessarily when members ran the new Kimi K3.1 Pro model. The model had launched only two days earlier, on October 24. As BAAI Community reported, the company took the blame directly, with one line saying the problem was on its side and should have been fixed sooner.

The compensation applies to every paid tier. Every member with an active subscription gets 25% of their monthly total back, which Moonshot says is roughly a week of use. Tencent News spelled out the mechanics: the share of allowance you have used drops by 25 percentage points, and anyone who had used less than 25% is reset to zero. The post said the reset was already done that morning.

The letter also promised fixes. Moonshot said it had changed its configuration and was cutting back on unnecessary sub-agent calls and on how long the model's answers run, so the same allowance covers more real work.

How one short prompt used up an allowance

Kimi plans work on a quota: a budget of usage that refills over time, including a window that resets every 5 hours. Tech星球, a Chinese tech outlet, documented two cases of that budget vanishing quickly.

In the first, a subscriber on the 199 yuan plan sent a single instruction to review a project's architecture. K3.1 Pro responded by starting 5 sub-agents, which are helper AI processes that split up a job and work in parallel. Within about a dozen minutes, the 5-hour window was exhausted.

In the second, the model was given the well-known test of drawing a pelican riding a bicycle. It generated for 1 hour 47 minutes and consumed 87% of a 5-hour allowance.

Tech星球 links this to how the model behaves by default. It reportedly thinks step by step in a long chain of reasoning for every request, handles very long inputs, and may call several sub-agents on top. A user sees one question, but the system does far more work behind it, and the allowance is charged for all of it.

Cheaper per token, not per task

A token is a small chunk of text, and AI companies price and meter usage by tokens. For developers who pay through Moonshot's API, the connection that lets software call the model, the price per million tokens fell by about a third. Input dropped from 20 yuan to 13 yuan and output from 100 yuan to 68 yuan. Artificial Analysis, an independent testing firm, lists first-party prices of $2 for input and $10 for output, which is 33% below the previous model, Kimi K3.

But Artificial Analysis also found that a typical task produces about 14% more output text than on K3. As a result, the average cost per task is about $1.98, roughly unchanged from K3. Tech星球 adds that running the firm's full Intelligence Index test produced about 72% more output than the median model tested.

The model scores well for the extra effort. Artificial Analysis gave K3.1 Pro 55 points on its Intelligence Index, fourth place worldwide and 11 points above K3's 44. It also scored 67% on Terminal-Bench 4.0, a test of command-line tasks, against 13% for K3. Moonshot says full model weights, the files that let others run the model, are due by November 4. If so, Artificial Analysis expects it to be the top-scoring open-weights model.

The limits did not disappear

The apology did not end the complaints. Tech星球 reports that users kept seeing a message saying Kimi was at peak hours and to try later or upgrade, even when they still had allowance left. Some of them had already upgraded once because their quota ran short.

The same outlet connects the pressure to other changes. On August 10, Kimi reportedly changed its memory feature without notice, adding a background process that reads conversations and spends paid quota. Recently, users reported memory not loading and older memory instructions being boiled down into general preferences. Tech星球 reads this as a possible attempt to save allowance.

Underneath, it points to computing capacity. After K3 launched, Kimi reportedly paused new subscriptions within 48 hours because its GPUs, the chips that run AI models, were close to full. Bloomberg is cited as saying Moonshot has a compute deal with Alibaba. Tech星球 sums up the squeeze: smarter models tend to cost disproportionately more to run, while subscription prices stay fixed.

As of November 3, the date of the English republication on KuCoin News, the allowance could be refilled, but the peak-hour message was still appearing.