An AI company constantly firefighting by resetting limits every day.
At the end of October, OpenAI adjusted its quota rules again in response to user complaints about reduced limits for the Pro 200 plan. After acknowledging “friendly feedback from users,” Tibo, head of Codex, announced that usage limits for all Pro and Business Premium plans have been doubled. The community responded with widespread celebration.
A similar incident recently occurred with Kimi, but the situation appears even more severe.
Less than 12 hours after the launch of Kimi K3.1 Pro, a new model under Moonshot AI, controversy quickly followed. On one side, overseas bloggers exposed that the model provided promotional templates requiring users to hide paid partnership labels during promotion; on the other, users voiced widespread complaints about rapid quota depletion—many found their usage limits nearly reached shortly after starting their tasks.
Two days later, Moonshot AI publicly apologized for the traffic throttling and unnecessary quota consumption related to the K3.1 Pro, restored 25% of the monthly quota as a one-time compensation for all subscription users, and stated that it would reduce unnecessary sub-agent calls and output lengths.
Although this is Kimi's third apology in six months, the "quota incident" has not truly been put to rest.
The "quota black hole" continues
Among China's large models, Kimi has always been a standout. With its strengths in long-text processing and text generation, it has not only built a loyal user base but also demonstrated the ability to compete head-to-head with overseas giants that possess vast computational power and financial resources.
But since the beginning of this year, complaints about its subscription plans being exhausted too quickly have never stopped. “Those without money can’t afford it, while those with money get better value from overseas models,” said a heavy AI user. At the same price point, Kimi offers far less credit than GPT and Claude. Other users have compared it to alternative models: “For 199 yuan, DeepSeek lasts me a whole month; with Kimi, it’s barely enough for one day.”
For these users, large models are not occasional chat tools but productivity products embedded into their daily workflows, making their consumption of quotas highly tangible.
After the launch of the K3.1 Pro, this contradiction became even more pronounced. The new model has seen significant upgrades in capabilities such as coding, long-range tasks, and 3D modeling, and has received a highly positive market response. Artificial Analysis assigned it an intelligence score of 55, ranking it fourth globally and making it the highest-scoring open-weight model on the list. Third-party evaluation agency ModelCompare even stated that, due to this model’s emergence, “the AI gap between China and the U.S. has shrunk from months to just 31 days.”
Looking solely at API pricing, Kimi continues to uphold the traditional cost advantage of domestic models: the K3.1 Pro reduces prices by about one-third compared to its predecessor, with input costs dropping from ¥20 to ¥13 per million tokens and output costs falling from ¥100 to ¥68 per million tokens. However, for users who purchase subscription plans, while model capabilities have improved, they must now calculate their usage quotas more carefully in practice.
A long-term subscriber of Kimi found that although the unit price is lower, the new model consumes more tokens per task, resulting in roughly the same overall cost. According to tests by Artificial Analysis, the K3.1 Pro generated approximately 150 million tokens across the entire Intelligence Index, while the median of the tested models produced only 87 million tokens. This means the K3.1 Pro’s token output is about 72% higher than the median level.
The real-world usage examples are even more striking. In the community, a user who subscribed to the Kimi 199-yuan plan sent just one command: “Review the project architecture.” K3.1 Pro immediately launched five sub-Agents, and within minutes, the entire five-hour quota was exhausted. When faced with the classic “pelican riding a bicycle” question, K3.1 Pro generated a response lasting 1 hour and 47 minutes, consuming 87% of its five-hour quota.
More troublingly, the quota compensation has not fully alleviated user dissatisfaction. Over the past few days, numerous users have reported that even with quota available, they are unable to use Kimi.
"Kimi is currently at peak capacity, please try again later or upgrade your membership for faster service." Even after paying for a subscription, this frustrating message continues to appear repeatedly in the chat boxes of many users—including those who have already upgraded their plan due to insufficient quotas.
Kimi has pushed the model to a higher capability limit, but what users now face is no longer just an issue of quota consumption—it's also the gap between peak service capacity and paid benefits.
“Hesitant to give fully,” is Kimi starting to access its memory?
To alleviate quota pressure, Kimi appears to have begun adjusting its once-prized memory feature. On August 10, Kimi updated its memory architecture without prior notice to users, introducing a "dream memory" mechanism that consumes paid quota by automatically reading conversation content and storing it in the memory database.
Several users have reported that Kimi’s memory space has been failing to load properly over the past few days, suggesting the official team may be upgrading and modifying the memory system. Currently, the new version tends to consolidate and summarize past memory instructions into general user preferences, rather than executing each memory entry individually as before. The benefit is potentially lower token consumption. However, the trade-off is a reduction in the previously more detailed memory capabilities.
However, this can only alleviate part of the consumption. Kimi's cost challenges primarily stem from the model itself.
The K3.1 Pro is a 3-trillion-parameter model. Previous industry information on the K3 series architecture suggested a model size of approximately 2.8T parameters and around 104B activated parameters. In comparison, models with relatively smaller parameter sizes and higher activation efficiency offer greater advantages in cost engineering control. Companies such as DeepSeek and Zhipu are optimizing model architectures and inference engineering to find a more balanced trade-off between capability and cost.
Secondly, there is the computational cost introduced by complex tasks. Kimi K3.1 Pro has thought mode enabled by default throughout the entire process, generating longer reasoning chains and more intermediate deductions. Long contexts increase the computational overhead of input processing, while complex reasoning and multiple agent calls further elevate total resource consumption. An agent task requiring multiple model invocations, execution of several sub-agents, and repeated review and revision may consume far more resources than users intuitively perceive from a single query.
For Kimi, what truly needs to be proven may not be whether it can continue to scale up model parameters, but rather its ability to improve reasoning efficiency alongside capability upgrades. A developer following Kimi suggests that instead of solely competing in the realm of high-end models, Kimi should perhaps invest more resources into developing lighter, more efficient models.
This is not an issue with Kimi alone.
All providers offering cutting-edge models on a fixed monthly fee face the same structural tension: improvements in model capability are often accompanied by proportional or even superlinear increases in inference costs, while subscription prices remain rigid and users expect that “what they pay for should be sufficient.” The difference lies in whether companies treat token efficiency as part of their product capability or treat breakthroughs in performance as brand assets—leaving subscribers caught in the middle, paying for runtime costs that exceed their expectations for “powerful intelligence.”
In addition, Moonshot’s computing resources are heavily dependent on external supply. Bloomberg previously reported that Moonshot has a computing agreement with Alibaba, granting access to a cluster of approximately 20,000 NVIDIA chips, which constitutes a significant portion of Moonshot’s overall computing capacity. Prior to this, within 48 hours of K3’s release, Kimi paused new subscriptions due to GPU capacity nearing its limit.
When a company’s own computing power is already strained, offering generous quotas to end-user subscribers is financially irrational.
Whom should domestic models be earning money from?
If a company has limited computing power, it must decide on a practical question: where should resources be prioritized—training and iterating on the next-generation model, investing in real-time inference services for millions of individual users, ensuring developer API usage demands are met, or providing subscribers with higher usage limits?
Different choices reflect varying commercial priorities. The controversy surrounding Kimi’s quota also reveals Moonshot’s true positioning for consumer subscriptions under computational constraints.
From August to September 2026, multiple media outlets reported that Moonshot AI was in negotiations with Microsoft, Amazon, and Google to list K3 on Azure, AWS, and Google Cloud, with Moonshot planning to take up to a 30% revenue share from sales of K3 services on these platforms. Additionally, K3’s open-source license is not the standard MIT license, but rather a custom agreement by Moonshot AI: MaaS hosting providers with annual revenue exceeding $20 million must sign a separate commercial license agreement.
Putting these two things together reveals that Moonshot is exploring a business path distinct from solely operating a consumer subscription model: rather than building large-scale consumer-facing inference infrastructure, it is far more aligned with its resource situation to “wholesale” its model capabilities first to cloud providers with computing power and distribution channels, rather than “retailing” them to millions of individual subscribers.
When the C-end subscription takes on more of a role as a user entry point, product feedback channel, and model capability showcase, rather than the primary source of revenue, its quota pool may differ from user expectations.
Currently, domestic large model companies are also seeking more suitable ways to generate revenue.
DeepSeek’s competitiveness stems largely from the combination of model capabilities, inference efficiency, and an open ecosystem. Currently, nearly all revenue comes from charges for model API calls, while its chatbot app remains free and ad-free. According to The Information, DeepSeek’s annualized revenue run rate has surpassed $1 billion, more than doubling from under $500 million just months ago. CEO Liang Wenhong explicitly stated at a recent investor meeting that the price increase for APIs has not led to customer churn, and user demand remains strong.
Zhipu has undergone a business model transformation, shifting from reliance on on-premises deployment to API subscription services. The first half of 2026 financial report shows that revenue from Zhipu’s open platform and API services accounted for 86.5%, a 2,735.7% year-over-year increase, while revenue from on-premises deployment dropped from 84.8% to 13.5%. This shift has been significantly driven by the surge in coding-related demand.
MiniMax, which began as a consumer-facing product for overseas markets, is seeking to expand its commercial boundaries. In addition to continuously developing models, APIs, and products for individual users, it is also intensifying efforts to penetrate the B2B market, extending its capabilities into video generation, as well as Agent, Reasoning, and Coding technologies. Its strategy is to integrate its model capabilities into a broader range of enterprise and developer scenarios, reducing reliance on a single product or customer base.
Compared to the direct monetization paths of model companies, internet giants have another advantage: they already possess mature product portfolios, large user bases, and enterprise customer resources, allowing them to integrate large models into their existing ecosystems and generate revenue through product enhancements and enterprise services. For example, Tencent is attempting to transform its products and office scenarios using large models. Products such as WeChat and Enterprise WeChat have gradually integrated capabilities from models like Yuanbao, Hunyuan, and DeepSeek, while AI office agents like WorkBuddy are rapidly expanding their user base.
After integrating the capabilities of QoderWork, MuleRun, WuKong, and other products, Alibaba launched Qwen Office, a unified enterprise-grade general Agent platform. At the Cloud Summit in September, DingTalk completed its commercial upgrade, introducing four paid tiers—AI Basic, AI Advanced, AI Premium, and AI Flagship—while retaining the free standard version, and fully integrating Qwen Office capabilities.
Low prices and traffic alone cannot sustain long-term business; model capabilities can continuously improve, but users’ willingness to pay won’t grow indefinitely. The real question is whether each company can find a business model aligned with its technological capabilities, computing resources, and market position.
Kimi is not the only one standing on the edge of a cost cliff.
Twitter: https://twitter.com/BitpushNewsCN
BitPush Telegram community: https://t.me/BitPushCommunity
BitPush TG subscription: https://t.me/bitpush