The next major challenge for large-scale models is commercialization.

The most renowned 'price cutter' in the large model sector is gearing up to raise its prices.

On August 6, DeepSeek posted an announcement in the user backend, revealing its plan to substantially increase the pricing of DeepSeek API services soon. Users were specifically cautioned that the 'expected price hike is significant' and were advised to plan their usage and recharge amounts wisely.

Image Source: DeepSeek

This news quickly garnered widespread attention. After all, over the past two years, whenever there has been a Token price war, DeepSeek has almost always been at the forefront. Now, DeepSeek, the pioneer in slashing Token prices, is proactively announcing a 'significant price increase,' giving the impression that the price war's initiator is preparing to bow out.

Coincidentally, on August 7, Reuters reported that Alibaba plans to implement a revenue-sharing model for large commercial users of its next-generation open-source AI large model, QianWen. Previously, only developers utilizing the model on Alibaba Cloud's platform were charged, in line with Alibaba Cloud's promoted MaaS (Model as a Service) model. However, if customers deployed the open-source model in their own data centers, they could use it entirely free of charge. Judging by the revenue-sharing model adopted by open weight models like Kimi K3, it appears that open-source large models will soon no longer be a 'free lunch.'

So, is the Token price war in the large model industry, which has persisted for over two years, truly nearing its end?

Before delving into the price hike, let's first examine how affordable DeepSeek currently is.

The official DeepSeek API page indicates that the current price for DeepSeek-V4-Flash cached input is 0.02 yuan per million Tokens, 1 yuan for non-cached input, and 2 yuan for output. The more powerful DeepSeek-V4-Pro costs 0.025 yuan, 3 yuan, and 6 yuan, respectively.

Image Source: DeepSeek

DeepSeek previously announced a peak-valley pricing mechanism, where prices for all billable items would double during peak hours from 9 AM to 12 PM and 2 PM to 6 PM Beijing Time. However, as of August 6, this mechanism had not yet been officially implemented.

Even so, considering the standard prices alone, DeepSeek-V4-Flash remains exceptionally affordable. Test data from overseas research institution Artificial Analysis reveals that the average cost for DeepSeek-V4-Flash to complete a round of comprehensive benchmark tests is merely about $0.03.

In comparison, the average cost for Kimi K3 to complete the same tests is about $0.86, OpenAI GPT-5.6 Sol about $1.86, and Anthropic Claude Fable 5 a staggering $3.15. In other words, under this test standard, the task cost of DeepSeek-V4-Flash is less than one percent of some overseas flagship models.

Image Source: Artificial Analysis

Of course, some might argue that the model capabilities differ, which is true. From Artificial Analysis's test rankings, DeepSeek V4 Flash 0731 Max scores 11 points less than Fable 5. However, when the price disparity widens to a hundredfold, the situation changes. After all, most tasks do not necessitate such high model performance.

Moreover, from the rankings, DeepSeek's prices are also 90% lower than Gemini 3.6 Flash High, which has similar scores. For most users, DeepSeek can truly be described as 'high quality and low price.' Therefore, DeepSeek's emphasis on the 'significant expected increase' naturally makes many developers apprehensive.

Especially in the Agent era, the tasks AI needs to perform are far from simple question-and-answer interactions. Take programming tasks as an example: to complete a task, Agent AI may need to read the entire codebase, generate a plan, call tools, run tests, and if errors are found, revise and call the model again for verification.

Image Source: Leikeji

Although it may seem like the user only inputs a single sentence, by the time the AI completes the task and provides a result, dozens or even hundreds of model calls may have occurred. In Agent mode, AI's Token consumption grows exponentially. When billions or even trillions of Tokens are consumed daily, even a slight increase in Token prices can lead to bills skyrocketing by hundreds of thousands or even millions of yuan.

To be honest, DeepSeek may not have anticipated this issue initially but only realized it after user usage surged. Simply put, DeepSeek may have expected its model to be popular but underestimated Agent's Token consumption, leading to rapid depletion of server computing power.

Moreover, according to Artificial Analysis's tests, DeepSeek V4 Flash 0731 Max is quite a 'chatterbox.' While other models complete tests with tens of millions to 100 million Tokens, DeepSeek uses 210 million Tokens. However, if DeepSeek's Token output could be reduced to a level similar to other models, costs could drop by another 50%, which would be daunting for other AI companies.

For most people, the biggest question is: Why is DeepSeek suddenly raising prices now?

The answer lies in the recently updated DeepSeek V4 Flash 0731 Max, which is incredibly powerful. Although its version number suggests it is inferior to V4 Pro, its Agent capabilities have significantly improved, which is exactly what heavy AI users need most currently.

This is reflected in the actual data, with a surge in call volume. According to the latest statistics from OpenRouter, DeepSeek-V4-Flash handled approximately 6.6 trillion Tokens in the past week, a 10% decrease from the previous week but still ranking first in the current model call rankings.

Image Source:

If it were just ranking first, it would be fine. What's even more astonishing is that among all models statistics by the OpenRouter platform, DeepSeek-V4-Flash also ranks first in usage, accounting for approximately 58% of the Token share. Data shows that from June 12 to August 6, the model handled approximately 97 trillion Tokens and completed over 10 million sessions.

Even though this is just third-party platform data, it is still surprising because DeepSeek stands out among top AI models like Gemini, ChatGPT, and Claude to capture this share.

However, this also highlights the computing power pressure DeepSeek faces, with trillions of Tokens processed weekly, which is indeed enormous for a quantitative enterprise. Previously, there were reports that DeepSeek owns a data center with a scale of 100,000 cards, which sounds substantial but is not on the same scale as AI companies like Google and OpenAI, let alone cloud service giants like Amazon and Alibaba Cloud.

In fact, since last week, there have been frequent reports of DeepSeek's computing power shortage and restricted usage. A few days ago, when I opened DeepSeek's web version to use the professional mode, the page directly indicated 'service busy.' To be honest, except for when DeepSeek first became popular early last year, I haven't seen this message on DeepSeek for a long time.

When DeepSeek first released its V4 version, it once announced future peak-valley pricing for Token services due to high call volume. However, this plan has not been implemented yet because DeepSeek found that its servers rarely reached 'valley' levels.

Online, DeepSeek users have shared official fault announcements, with two performance degradation notices issued during working hours in China and working hours in the United States (Chinese time). At night, Chinese users indeed stop using it, but American users start 'getting up and pedaling,' so where is the 'valley'? It's all 'peaks.'"

"For DeepSeek, the current pricing barely allows it to break even, let alone make a profit. Many industry insiders speculate that DeepSeek has been subsidizing Token prices; otherwise, given its current computing power scale, Token prices could not be so low (hence, many netizens refer to Liang Wenfeng as 'Saint Liang,' thanking him for open-sourcing and ultra-low-priced Tokens).

Image Source: Weibo

Of course, there's no way to prove this unless DeepSeek admits it. However, one thing is certain: while low prices attract developers, excessively low prices bring another problem: all sorts of miscellaneous demands flood toward DeepSeek.

For example, a developer is writing a program using AI and encounters a small bug. He could spend ten minutes debugging it himself, but because DeepSeek's prices are so cheap, running a debug round might only cost a few mao, so he chooses to let DeepSeek handle it instead.

Conversely, if he were using a high-priced model like Claude, where running a task costs several or even tens of yuan, he might not want to let the AI run it and would choose to debug manually instead. Do not underestimate the difference between these two mindsets. When most users behave this way, DeepSeek actually uses a significant amount of computing power on very simple tasks.

Moreover, these requests also occupy computing power shares for complex tasks, leading to the 'congestion' mentioned earlier. Of course, since everyone has paid, how they use it doesn't matter, but for DeepSeek, they do need to consider users who use Tokens for commercial projects more and ensure their experience as much as possible.

Therefore, after adhering to low prices for a while, DeepSeek's price hike was within many people's expectations. Moreover, to be honest, when DeepSeek announced its price cut, it used the term 'limited-time,' suggesting it had long prepared for a price increase.

From a consumer's perspective, I naturally hope DeepSeek can continue offering discounts, but from an industry perspective, relying on subsidies to preempt market share is unsustainable, especially in an industry like AI, which requires heavy investment and rapid iteration. Companies must consider breaking even.

Moreover, DeepSeek's price advantage is so significant that even with severalfold price increases, it still offers excellent cost-effectiveness and remains the most cost-effective choice for most developers.

Although some netizens have suggested that DeepSeek's price hike signals the end of the Token price war after its announcement, I do not believe so. Instead, I think the price war has just begun.

Almost at the same time,

The price war over the past two years has been quite straightforward: manufacturers competed based on the price per million Tokens. Today, you announce a drop to 5 yuan, and tomorrow, your competitor can offer 2.5 yuan to challenge you. Some AI companies even distributed free quotas in the millions of Tokens, followed by limited-time discounts and caching discounts, causing Token prices to drop to a level where developers could almost ignore them.

However, as model capabilities and application scenarios continue to evolve, comparing Token prices alone no longer makes much sense. First, different models consume different numbers of Tokens to complete the same task. Taking Artificial Analysis's test projects as an example, DeepSeek used 210 million Tokens, while Gemini 3.6 Flash High used 59 million Tokens, a gap of nearly 3.5 times.

Image Source: Artificial Analysis

This is why Gemini 3.6 Flash High's Token output price is almost 20 times that of DeepSeek, yet the cost difference for their test tasks is only 10 times. As task complexity increases, the price difference may further narrow, affecting model competitiveness.

Second, model speed and stability also begin to impact enterprise costs. If your API responds slowly, frequently times out, or is unavailable during peak hours, it causes severe experience issues. Especially for businesses requiring real-time calculations and high stability (such as financial services), stability becomes even more critical.

As AI truly integrates into the underlying operational chains of various industries, a noticeable change is that, beyond price, enterprises also have more requirements for computing power: stability, performance, and speed. For DeepSeek, although it does not rely on Tokens to make money, if it wants to continue occupying a place in the model field, it must consider B-side demands.

At this moment, the most suitable approach is to "dissuade" some users through appropriate price increases while enhancing the experience of high-value users. Although this may sound a bit harsh, from a corporate perspective, it is quite understandable. Moreover, given that V4 Flash 0731 has already demonstrated such performance, subsequent versions like V4 Pro xxxx might be even more powerful. If the price gap becomes too large at this point, it would also make it difficult for DeepSeek to balance the cost-effectiveness between the two versions.

In fact, when looking at the global AI model market, it's not hard to notice that everyone has started to "segment" their models in recent months, establishing a tiered performance and pricing system to meet different user needs. Therefore, rather than saying the price war has ended, it's more accurate to say that it's only just beginning. AI companies are now entering a true "hand-to-hand combat" phase, hoping to capture more B-end market share by offering more cost-effective models and more stable, faster services.

It can be said that the commercialization battle for Tokens has only just officially begun.

DeepSeek QianWen Open-Source Large Model Alibaba

Source: Leikeji

Images in this article are from: 123RF Royalty-Free Image Library       Source: Leikeji