DeepSeek is protecting compute for paying API developers, dropping their server-side search fees on November 18 while free consumer users lose capacity.
DeepSeek has started rationing the free version of its chatbot, according to user reports gathered by Yicai and other Chinese outlets. People using the app and web chat now run into a daily quota, signaled by a popup that reads "因服务器繁忙,使用额度已达上限" (the usage quota has been reached because the servers are busy). Some users hit the ceiling after less than two hours of continuous use. A single text input is now capped at about 40,000 characters, and long conversations that sit idle for too long can no longer be continued. The only reason users were given is insufficient computing power.
The Hangzhou company has not published unified quota rules, explained the change or said when the limits will end, Yicai reported. It also offers no paid consumer plan that would let heavy users buy their way past the caps. The application programming interface (API), through which developers pay for model access by usage, is outside the new limits and keeps its own usage limits and pay-as-you-go billing.
The rationing follows a surge in demand that DeepSeek’s capacity has struggled to absorb. DeepSeek’s V4.1 Flash model processed 59.2 trillion tokens on OpenRouter in the week to November 15, the most of any model worldwide. A 36Kr piece, republished from Tech商业, put the comparable figure for the week of September 14 at 15.8 trillion.
Reports differ on when the limits started and how strict they are
The accounts agree on the daily quota and the 40,000-character input cap but diverge on the details. Caixin Global says the restrictions began on November 15 and that long conversations inactive for 72 hours can no longer be processed, which the company attributed in user notices to "a shortage of computing resources." Yicai and 36Kr date the first user reports to November 16. Both say conversations idle for more than seven days are no longer supported, while some users were cut off after only 72 hours.
Caixin also reports a cap on document reading: the chatbot takes in only the first 20 pages of a document, or 10,000 tokens. Users who reach a cutoff are prompted to start a new conversation or export their chat history, according to Yicai. Yicai Global noted that light users may barely notice, but people who use the chatbot for long stretches, such as revising papers, organizing long documents or multi-day role-play, are hit hardest.
Paying API customers come first as the developer business grows
An industry insider quoted by Yicai Global said DeepSeek may have to balance compute between model training, its paid API and free chat, and that capping input length, idle conversations and daily use all reduce token consumption. Yao Shuqiao of Haitong International told Caixin that DeepSeek has fewer computing resources than the major cloud providers, so it must allocate them efficiently and prioritize paying customers. He expects computing power to stay tight in the short and medium term.
The API side is moving the other way. From 12:00 Beijing time on November 18, server-side web search and web fetching triggered during model calls are no longer billed separately, and only model token usage is charged, according to DeepSeek’s pricing page and IT Home. The standalone Search and Fetch APIs still cost US$0.75 per 1,000 successful requests. The change comes days after DeepSeek opened paid application APIs and tools on November 13 that let developers rebuild much of its chatbot’s feature set.
Capacity has been under strain for weeks. 36Kr counted at least five service outages since October, including one of about five hours on October 15. It also cited reports that DeepSeek closed a second financing round of about RMB105.5 billion in October at a valuation of about RMB615.5 billion, with the money mainly earmarked for compute infrastructure.
A paid consumer tier is the change users are asking for
Users have complained on Xiaohongshu and offered to pay for a subscription, Caixin reported, and Yicai said suggestions range from monthly to annual plans. A developer found quota-management logic and references to a "coding plan" in the front-end code of DeepSeek’s web version, according to Yicai. 36Kr said DeepSeek had not commented on those findings.
Other Chinese labs have run into similar walls. Caixin noted that Moonshot AI suspended new Kimi K3 subscriptions in July and later apologized for rate limits on K3.1 Pro. For DeepSeek, the signals to watch are published quota rules, an end date for the limits, or a priced consumer plan, any of which would show how the company intends to split capacity between free users and paying developers as it adds compute.