A Moonshot AI, the Chinese developer of the Kimi artificial intelligence assistant, announced compensation to subscribers after acknowledging problems with usage limits and unnecessary consumption of quotas for Kimi K3.1 Pro. The company promised a one-time replenishment equivalent to 25% of the monthly quota for users on all plans.
The measure was announced by Kimi's official profile on October 25 and reported by Reuters this Monday, the 26th. According to the company, the compensation is equivalent to approximately one week of use. For those who had consumed less than 25% of their quota, the usage record would be reset.
The compensation is not a cash refund. It is a restoration of usage capacity within the subscription. Moonshot said that the consumption indicator in each user's account would reflect the adjustment by the end of the day in U.S. Pacific Time. No independent audit confirming the measure's application to all accounts was disclosed.
Unnecessary agent calls increased consumption
The company said it had adjusted K3.1 Pro parameters to reduce unnecessary subagent calls and the length of responses. In AI systems capable of delegating tasks, a user request can trigger additional operations, which also consume computational resources and the contracted quota.
The explanation offers an important detail about the incident: part of the consumption was not necessarily associated with additional requests made directly by subscribers, but with the operation of the system itself. The company acknowledged responsibility for the usage limits and for the spending considered unnecessary and apologized to customers.
The fix was announced days after the launch of Kimi K3.1 Pro. Reuters noted that Moonshot had already faced computational capacity difficulties in a previous version of the model. This time, the response involves both a temporary replenishment of the quota and promised adjustments to the service's behavior.
Moonshot did not say how many subscribers were affected or present metrics that would allow measuring how much the changes will reduce consumption in future tasks. The practical effect will depend on the execution of the fixes and on the agents' behavior during real-world use.
