Kimi apologizes for K3.1 Pro throttling and quota drain, credits subscribers 25% of monthly allowance

Early on October 26, Moonshot AI's Kimi assistant published an apology to paying members over throttling and unnecessary quota consumption they encountered while using the K3.1 Pro model, and attached a compensation plan. According to Jiemian News, the notice opens with a direct apology to users before laying out the credit rules.

Kimi said it would run a one-time reset for subscribers at every tier that day: all members within their subscription period receive a credit equal to 25% of their monthly quota, roughly a week's worth of usage, and accounts whose current usage is below that 25% mark are reset to zero. The usage bar is supposed to reflect the adjustment by the end of the day.

Alongside the credit, Kimi described engineering fixes: configuration has been adjusted, and the company is reducing unnecessary sub-agent calls and output length so that the same quota covers more real work. That is an acknowledgment that the drain was not all genuine user demand, but partly overhead generated by how the system orchestrates tasks.

For a subscription AI product, the quota is the pricing. Members pay monthly for a predictable amount of usage, so throttling and phantom consumption damage more than a single session, they weaken the basic expectation of what was purchased. A flat 25% credit converts losses that are hard to audit case by case into a standard that is calculable and publicly stated.

The scale of the remedy is not trivial. A quarter of a monthly allowance is about a week of usage, and it covers every active tier; the rule that balances under 25% are zeroed out returns light users to a full tank. For the platform, it is a spend of quota in exchange for stopping reputational damage.

Notably, Kimi points to sub-agent calls and output length as the cause. That kind of overhead is common in agentic products: to finish a task, the model decomposes steps, calls tools repeatedly and generates long intermediate results, which inflates the cost of a single request while the user only sees the quota bar falling fast.

What to watch next is whether the fixes actually land: whether average consumption per task comes down, whether throttling disappears at peak hours, and whether Kimi publishes more granular usage accounting so members know where their next deduction went.