DeepSeek limit, confidence comes from outside the chat box

2026-11-21 10:08

Author Ai Lingxin

On November 16th, many DeepSeek users found that free chatting had been tightened: long conversations that were inactive for more than 7 days could not be continued, the maximum input limit was lowered to around 40000 words per session, and daily conversations were also set with a limit. When used up, they would prompt "the limit has been reached". Some users have expressed willingness to spend money to purchase more credit, but DeepSeek does not provide a personal membership entrance.

On November 13th, DeepSeek announced the opening of a new batch of paid APIs (interfaces for other programs to call), providing developers with network search, file processing, voice and other functions. Other companies can even use it to create products similar to the DeepSeek App.

DeepSeek's chat box is limited, while the service for others to create applications is expanding. With billions of users, why isn't it in a hurry to sell members and instead willing to open up its technology to others?

Putting research and development first

The direct reason for the limit is the tight computing power. In May of this year, Liang Wenfeng admitted at a closed door investor exchange meeting, "What really limits us is resources, especially computing power resources

According to OpenRouter, a large model API aggregation platform, the call volume of DeepSeek V4.1 Flash reached 59.2 trillion tokens from November 9th to 15th, ranking first among all models in the world and about four times that of two months ago. According to the official status page of DeepSeek, there have been multiple service interruptions or anomalies on the platform since October.

Not enough computing power, why compress for free chat first? Free chat, paid API interfaces, and model training all consume computing power, and companies must make trade-offs in computing power allocation. The non profitable option will naturally be compressed first. But not making money is just one of the reasons. Who uses computing power first depends mainly on what the company ranks first.

Since its inception, DeepSeek has aimed at General Artificial Intelligence (AGI), an artificial intelligence system that can reach or exceed human level in almost all cognitive tasks. In May 2023, Liang Wenfeng said in an interview with 36Kr's "Dark Surge" that "our goal is also very clear, which is not to do vertical classification and application, but to do research and exploration". When asked about the same topic again in 2024, his answer remained unchanged: "Research and technological innovation will always be our top priority." In May of this year, Liang Wenfeng said at a closed door investor exchange meeting that he had no intention of turning DeepSeek into the next ByteDance or Tencent. Both the C-end and B-end businesses are just "byproducts" of DeepSeek's AGI journey.

Free chat does not charge users, so it seems natural to launch membership. DeepSeek may not necessarily not want to do this business. Just a few days before the limit, a developer discovered a complete limit management function in the front-end code of its web version, speculating that the company is preparing a personal payment plan. But it has not sold members for a long time, and the main reason may be that the current computing power does not allow it. Once a member is sold, the platform must guarantee a certain amount according to the agreement, and the originally flexible free usage has become a promise that must be fulfilled. To expand computing power, both equipment procurement and construction require time.

DeepSeek has a reason for choosing a limit, but for users who have been using free chat for a long time, it is more friendly to explain the rules in advance. If membership development is really on the way, we also hope that DeepSeek can provide a better answer for users who trust it.

Confidence comes from API commercialization

DeepSeek prioritizes the development of basic models and strives for more model businesses behind applications. Liang Wenfeng described this division of labor in 2024: 'We are only responsible for basic models and cutting-edge innovation, while other companies build their toB and toC businesses based on DeepSeek.'. ”

In other words, even if users have not opened DeepSeek, as long as the product they use is connected to DeepSeek's official interface, it can charge based on usage.

Why are there still people willing to purchase the official API when the model is already open source and can be downloaded for free by enterprises? Because companies deploy DeepSeek models themselves, they also need to purchase servers equipped with AI chips or rent cloud computing power, and then arrange technical personnel to maintain the system. For developers who lack equipment and technical teams, purchasing official services directly is cheaper and more convenient.

The API price of DeepSeek has always been relatively low among top model manufacturers. In the 2024 domestic price war for large models, Liang Wenfeng put his principles very bluntly: "Do not subsidize, nor make huge profits." At a closed door investor exchange meeting, he introduced that API pricing is based on the cost recovery of newly purchased equipment in about ten months, and "I think this is a reasonable profit.

DeepSeek dared to lower its quote because the new model structure first lowered the cost. In May of this year, DeepSeek permanently lowered the API price of its flagship model to a quarter of its original price. After being hit by the input cache, the minimum cost per million words is only two and a half cents. Third party calculations show that under the same task complexity, its average call cost is about one tenth of GPT-5.5.

The cost of using the model is continuously decreasing. According to the Economic Observer on October 22, a user who used the DeepSeek open-source intelligent agent application Harness to analyze literature presented two bills before and after the model update: the number of requests increased by about 13%, the consumption of keywords increased nearly threefold, and the cost slightly decreased, mainly due to an increase in cache hit rate and a decrease in output keywords - caching allows the system to reuse processed information and reduce redundant calculations.

Although the pricing is low, the revenue scale of this business is already quite considerable. On November 5th, the DeepSeek Harness team revealed that, based on users using the official API, the daily active users reached 1.54 million on November 4th, with a daily keyword consumption of approximately 92 trillion. Based on the daily call volume, cache usage, and API prices, Caijing estimates that if this level continues for a month, the corresponding monthly revenue would be approximately 192 million yuan.

Liang Wenfeng spoke more directly at the exchange meeting about how confident he really is. He predicts that if demand continues to expand and more graphics cards can be purchased, API revenue is likely to reach several billion dollars. Even if technology no longer advances, 'in the worst case scenario, selling APIs may be able to support a listed company... selling APIs to the fullest extent and doing these services well, I think it's enough'.

(The author of this article is a researcher at the Jingguan Research Institute)