Moonshot AI published a post in the early hours of October 26 apologizing for the rate limiting and excessive quota consumption that members ran into while using Kimi K3.1 Pro, and said it is resetting quotas for subscribers on every tier. The model went live in the early hours of October 24, so only two days have passed.
The apology is blunt:
"We apologize for the rate limiting and unnecessary quota consumption that Kimi members have experienced when using K3.1 Pro. The problem was on our side, and we should have fixed it sooner."
How the make-good works
There is a single rule. Members whose subscriptions are active receive compensation equal to 25% of their monthly total, which the company converts to roughly a week of usage. Anyone whose current usage is below 25% of the monthly total gets their usage reset to zero. Usage meters will be updated before the end of October 26.
On the technical side, the company says it has changed its configuration and is reducing unnecessary sub-agent calls and output length so that the same quota gets more real work done. That amounts to an admission that fast quota burn was partly the model overthinking: splitting a single request into too many subtasks and writing overly long answers both draw down a member's quota by token count.
The statement does not say when the rate limiting began, how many members were affected, or which tiers were hit hardest.
What K3.1 Pro is
K3.1 Pro is the first model in the K3.1 series. The company calls it its "most capable model to date at the 3-trillion-parameter scale," with a 1 million token context window. A faster Highspeed variant launched alongside it, and both are available on kimi.com, the Kimi app, Kimi Work, Kimi Code and the API. Full weights are planned for release by November 4.
According to a rundown by Huxiu, K3.1 Pro still has 2.8 trillion total parameters, the same as K3, which was released on July 16. It was not pretrained again; the gains come mainly from more post-training compute, with emphasis on coding, 3D modeling, financial research and long-horizon agent tasks. A 3D Make feature launched at the same time, letting users preview and edit models online and export files for 3D printing. The company's own benchmarks compare the model with Claude Opus 5.5 and GPT-6 Astra, and its reported Terminal-Bench 4.0 score is 65.7%.
On the composite intelligence index from third-party firm Artificial Analysis, K3.1 Pro scores about 55 against about 44 for K3. The improvement comes mostly from agentic tasks, while gains on some knowledge and scientific reasoning tests are far more modest.
Set against China's Flash tier
API pricing is $2 per million input tokens and $10 per million output tokens, about a third lower than K3. Next to lightweight Chinese models from the same period, the gap is still large:
| Model | Intelligence index (approx.) | Input / output (USD per million tokens) |
|---|---|---|
| Kimi K3.1 Pro | 55 | 2 / 10 |
| DeepSeek V4.1 Flash-1017 | 45 | Same as MiniMax off-peak, double at peak |
| MiniMax M3.1 Flash | 43 | 0.15 / 0.60 |
A rough calculation, with 1 million input and 1 million output tokens each: Kimi costs $12, MiniMax and off-peak DeepSeek cost $0.75, and peak-hour DeepSeek costs $1.50. K3.1 Pro is positioned as a flagship, charging more than ten times the price for an index lead of about ten points.
The situation on the membership side is more delicate. API usage is metered, so you pay for what you use; subscribers get a fixed quota, and the more eagerly the model calls sub-agents, the faster that quota burns. Shortly after K3 launched in July, Kimi paused new consumer subscriptions because of tight compute. This time the problem has shown up in a different form, in the quotas. Kimi says it is developing a Flash version, but has not said when it will arrive or whether it can ease the compute pressure on the membership side.
Sources: Huxiu, The Paper, CocoLoop, National Business Daily; the compensation ratio and reset rules were checked against Kimi's official WeChat account, while intelligence index figures and API prices should be read against Artificial Analysis and each vendor's official pricing.