Less than a week after launch, Kimi K3.1 Pro has pushed subscriber quotas into the spotlight. On 26 October 2026, Moonshot AI apologised through Kimi's X account, admitting that the new model's launch brought rate limiting and unnecessary quota consumption, and announced it would cut every subscriber's used monthly quota by 25 percentage points. Reuters reported the statement the same day.
What the 25% actually means
According to Moonshot, 25 percentage points is roughly one week's usage. Subscribers who had used less than 25% have their usage reset to zero; those who had used more get a real extra week of quota. This is not a new month, nor a full refund, and the Reuters report mentions no refund. What subscribers should check is whether their used-quota figure has actually gone down, not just whether an apology arrived.
Subagents are the part burning the quota
Moonshot also said it has adjusted parameters to reduce unnecessary subagent calls and output length. Subagents are helpers a main agent sends off to do pieces of a job: searching, running code, checking results, each extra dispatch burning more tokens. Kimi K3.1 Pro is built for long tasks, and the longer a task runs, the easier it is for dispatched calls to spiral. Whether tighter parameters can save quota without noticeably slowing tasks is the real thing to verify after this compensation, and the report gives no figure for it.
This is not an isolated launch accident
Moonshot already ran into computing-capacity strain after the Kimi K3 launch, and the same kind of problem has now recurred with the new model. Days earlier, the same company deliberately curbed Kimi K3.1 Pro's cyber-offensive capability, removing cyber-related data from its training corpus and reinforcement-learning environments. Taken together, model capability has moved ahead while the product, compute and safety margins around it have not kept pace. Reuters also points to a wider backdrop: Chinese AI labs face computing shortages under US chip export controls.
What subscribers should watch next
Three things matter: whether the rebate lands automatically, since the report gives no timetable; whether quota burn on long tasks genuinely slows down; and whether Moonshot will warn users before any future rate limiting instead of compensating afterwards. An apology and a rebate can buy back trust once. If subagent calls and output length are not fixed, the next big model launch will bring the same quota dispute back.
