Moonshot AI posted a public apology on Sunday to subscribers of its new flagship model, admitting that Kimi K3.1 Pro had been burning through paid quotas faster than it should. "We're sorry for the rate limits and unnecessary quota consumption on the K3.1 Pro subscription plan," the company wrote. "These were our issues, and we should have fixed them sooner." Every subscriber on every plan receives a one-time reset restoring 25 percent of their monthly quota, which the company estimates at roughly a week of usage, and users who had consumed less than 25 percent are reset to zero.
The interesting part is the diagnosis. Rather than cutting prices or raising limits, Moonshot said it is "reducing unnecessary subagent calls and output length, so your quota goes further." A subscription to an agentic model is effectively a bet on how many tokens the orchestration layer will spend on your behalf, and Moonshot just conceded that its own agent loop was spawning too many subagents and writing too much. Independent tracking by Artificial Analysis supports the mechanism: Kimi K3.1 Pro costs about $1.98 per Intelligence Index task, roughly the same as its predecessor K3, because average output per task grew from around 48,000 to about 55,000 tokens, a 14 percent increase that ate the per-token discount.
K3.1 Pro launched late last week as a 3-trillion-parameter-class open-weight model with a 1 million-token context window and an input price of about $2 per million tokens, with full weights promised by November 4. Moonshot has claimed the model beats OpenAI's GPT-6 Astra on coding-agent work and approaches Claude Opus 5.5, including a jump from 13 percent to 67 percent on Terminal-Bench 4.0, figures that remain company self-reports until the weights land and community evaluation begins.
The apology arrives two days after launch, which is fast even by the standards of Chinese consumer AI. The precedent weighed on the company: in July, the K3 launch drew 48 hours of demand far beyond forecasts, forced a pause on new consumer subscriptions, and turned capacity planning into a public story. Moonshot's consumer business has since become a real revenue line, with Bloomberg reporting an annualized run-rate above $1 billion across Kimi Work, Kimi Code and developer API sales as of August, against a year-end target of $2 billion.
What the apology conspicuously did not address is the other K3.1 Pro story. VentureBeat reported over the weekend that a promotion briefing document instructed influencers to post about the model without the paid-partnership label, and to frame it as an alternative to Claude Opus 5.5. The document has not been independently authenticated, and it remains unclear whether it originated from Moonshot or an outside agency. Moonshot's statement covered rate limits and quota mechanics and said nothing about the brief. Until it does, the safe operating assumption is that launch-week enthusiasm for any model, including glowing demos on X, should be treated as potentially paid placement.
The episode also lands on a specific nerve in the subscription business. Users complained they could not predict how much quota a single task would consume, because the model decides on its own how many subagents to spin up and how long to write. That opacity is now the product's biggest reputational risk, bigger than any benchmark gap. Quota resets, meanwhile, are hardening into a routine retention tool across agent subscriptions; OpenAI pushed a similar reset for Codex the same weekend.
The second moment of reckoning comes on November 4, when the weights are due. If independent replication confirms the benchmark claims, technical credibility recovers and the influencer affair fades into a marketing footnote. If community testing surfaces gaps between the leaderboard and real-world coding work, the damage compounds, because the company will have paid influencers to amplify numbers its own users already distrusted on quota grounds.
Comments (0)
Log in to join the discussion
Log InNo comments yet