Kimi's flat-fee plans draw every agent task from one monthly pool, so Moonshot's subscription margins depend on how much compute each K3.1 Pro task burns.
Moonshot AI has handed every paying Kimi member a quarter of a month’s usage back, two days after launching its Kimi K3.1 Pro model. In a post published in the early hours of October 26, the Beijing company apologized to members for rate limiting and for quota that was consumed when it did not need to be while they used the new model, The Paper reported, reproducing the statement. Reuters, in a report carried by The Star, said the company made the same announcement on its Kimi account on X.
The compensation applies to subscribers on every plan whose membership is still valid. Each receives 25% of their monthly total, which Kimi says equals about one week of use; anyone who had used less than 25% of the month’s allowance is reset to zero. The adjustment was to appear in members’ usage bars before the end of October 26, according to Yicai and the 21st Century Business Herald. Kimi also wrote that the problems were its own fault and that it should have fixed them sooner, according to a second report in The Paper.
The more telling part of the statement is the fix. Kimi said it had adjusted its settings and was reducing unnecessary sub-agent calls, where the main AI agent hands pieces of a task to other agents, and shortening the model’s outputs, so that the same quota completes more real tasks. Moonshot did not say how much compute the changes save or how many members were affected.
K3.1 Pro’s agents ran through a quota pool shared by every Kimi feature
Moonshot announced K3.1 Pro on Saturday, October 24, describing it as its most capable open-weight model in the 3-trillion-parameter class, with a 1-million-token context window, according to City News Service. A larger model doing longer agent work draws more compute per request, and Kimi’s subscription design passes that cost straight into members’ allowances.
Kimi’s membership pricing page lists five current plans, from Go at RMB49 a month to Ultra at RMB1,399 a month. All membership features draw from one shared monthly pool that is deducted by actual use, so heavy use in one feature leaves less for the rest. Unused quota does not carry over. The page says Ultra includes 28x Agent credits, twice those of the Max plan, and up to four Agent tasks running in parallel. Some features charge the pool even when idle: Kimi Claw’s cloud host deducts 0.6% of the quota every day at 4 p.m.
Under that design, an agent that spawns extra sub-agents or writes longer answers than needed spends a member’s flat monthly fee faster without the member doing more. Moonshot’s settings change cuts that consumption, and so lowers its own compute cost per subscriber, while the 25% credit adds to that cost for the rest of the month.
Kimi launches in April, July and October all strained Moonshot’s capacity
Moonshot has had to ration or refund capacity after Kimi model launches in April, July and October. After Kimi K2.6 went live on April 20, members ran into queues and features that were briefly unavailable, and an error in the back-end tally of Agent quota deducted some users’ credits by mistake, IT Home reported. Moonshot then reset every user’s monthly quota to 100% on April 22.
In July, 48 hours after K3 launched, Moonshot stopped selling new subscriptions. "Demand has pushed close to the limits of our current capacity," the company said on X, according to The Decoder. It said it would direct its available computing resources to existing subscribers, TechNode reported in July, and it split its plans into a general Kimi membership and a separate Kimi Code membership to spread compute more evenly, The Decoder added.
October’s response was narrower than either. Kimi’s statement announced no pause in sales and no full quota reset; it changed how the model spends compute instead. Reuters linked the episode to wider strain, noting that Chinese AI labs face computing shortages under US chip export controls.
Quota terms will matter more as Moonshot courts paying users
Moonshot is building up its paid base through channels such as an Agricultural Bank of China promotion that ties Kimi membership to card spending. It is also preparing a Hong Kong listing: in September, TechNode reported that Moonshot had submitted a confidential application, which the company declined to comment on. How much work a member gets for a fixed fee shapes both subscriber growth and the margins investors would weigh.
Moonshot has not said whether its trims to sub-agent calls and output length will stay in place once demand for K3.1 Pro settles, or whether they change the quality of results members get. Shorter answers and fewer helper agents save compute, but members bought plans to finish larger tasks. The clearest signal will be the next revision of the pricing page: whether the Agent multipliers it lists for each plan rise, hold or shrink after the October 26 reset.