Author｜Xiang Qing, Editor｜Zhao Yuan

On August 6, DeepSeek made an announcement: 'We intend to substantially raise the prices of our DeepSeek API services in the near future. Please plan your usage accordingly. The specific details will be communicated through a formal notice.'

Over the past year, DeepSeek has been prominently labeled as a provider of 'low-cost' services. With API prices significantly lower than industry standards, DeepSeek quickly expanded its developer user base and propelled the global large-scale model industry into a price war.

The current price hike not only reflects pressures from computational costs but also, more significantly, marks a pivotal shift from technological innovation to commercialization.

According to Bloomberg, DeepSeek is gearing up for an A-share listing, aiming to complete its IPO by 2027, with a potential launch as soon as the end of this year. A new round of financing is also in progress, valuing the company at approximately $71 billion before investment.

Raising prices on the brink of its IPO serves as a testament to DeepSeek's previous successful path.

If users continue to utilize the service post-price increase, it signifies that DeepSeek has genuinely fostered user loyalty. Conversely, if users migrate to cheaper alternatives, it suggests that the competitive barriers built over the past year were primarily price-driven.

Prior to this comprehensive price hike, DeepSeek had already hinted at adjustments to its pricing strategy.

In mid-July, DeepSeek attempted to introduce a 'peak-valley pricing' mechanism, planning to double API prices during weekday peak hours (9:00-12:00, 14:00-18:00) to encourage non-urgent tasks to be scheduled during off-peak times and alleviate computational congestion.

This approach mirrors dynamic pricing in the cloud computing sector, where prices escalate during periods of high demand to mitigate resource congestion and revert to competitive levels when demand subsides. However, just three weeks later, DeepSeek transitioned from partial adjustments to an overall price increase.

This indicates that DeepSeek's challenges extend beyond resource constraints during peak periods; its entire business model necessitates restructuring.

Previously, DeepSeek's price advantage was remarkable.

In terms of current pricing, DeepSeek's API prices are differentiated for 'input tokens' and 'output tokens,' with variations based on 'cache hits' and 'cache misses.' For the V4-Flash model, input (cache hit) costs 0.02 yuan per million tokens, input (cache miss) costs 1 yuan per million tokens, and output costs 2 yuan per million tokens. For the V4-Pro model, input (cache hit) costs 0.025 yuan per million tokens, input (cache miss) costs 3 yuan per million tokens, and output costs 6 yuan per million tokens.

These price levels are among the lowest globally for mainstream models. Some analyses highlight that, while maintaining highly competitive benchmark performance, the V4 Flash is currently the most cost-effective well-known large-scale model to operate.

According to test data from Artificial Analysis, when multiple mainstream models executed the same benchmark, DeepSeek V4-Flash completed a task for a mere 3 cents. Anthropic's flagship Claude Fable 5 cost $3.15, and GPT-5.6 Sol cost $1.86.

Low prices have fueled rapid growth, validating DeepSeek's strategy of trading low prices for market expansion.

Overseas market data reveals that DeepSeek V4-Flash has emerged as one of the most widely adopted large-scale models globally. The latest weekly rankings (July 27 to August 2) from the global large-scale model aggregation platform OpenRouter show that DeepSeek V4-Flash ranked first with 7.22 trillion tokens in weekly usage, outpacing competing models like Xiaomi's MiMo-V2.5.

DeepSeek has demonstrated that high-performance AI capabilities need not be sold at exorbitant prices. It has not only compelled competitors to revise their pricing strategies but also redefined the competitive landscape in the large-scale model industry, creating a market scenario investors dub the 'DeepSeek Kill Zone.' Model vendors must either reduce prices to match DeepSeek's cost advantage or invest more resources to develop significantly superior models—otherwise, they face elimination.

However, the market advantage garnered through low prices also comes with substantial commercial pressures. Behind affordable AI lies a sustained increase in computational costs.

Low prices were never the ultimate objective but a means to acquire users and capture market share. Now that the user base has expanded, the question DeepSeek must address has evolved. The focus is no longer on attracting users but on retaining them and converting them into revenue.

This price adjustment is, in essence, a litmus test for DeepSeek's previous success model. Do users opt for DeepSeek because it is affordable, or because it is robust?

If users persist with the service post-price increase, it indicates that DeepSeek has established product value beyond price advantages, with users acknowledging its model capabilities, service stability, and ecosystem experience. However, if users swiftly migrate to cheaper alternatives, it suggests that DeepSeek's previously erected competitive barriers were more price-centric than rooted in genuine user loyalty.

This is the ultimate test of the price hike.

DeepSeek's price increase compels the entire industry to confront a reality: if even the most cost-effective large-scale model vendor is raising prices, the era of nearly free AI and user subsidies may be nearing its end.

The rules of the AI game are evolving. Previously, the emphasis was on who could offer the lowest prices and acquire users the fastest. Now, the challenge lies in achieving commercialization while managing costs.

This year, several domestic large-scale model companies, including Zhipu, Yuezhia'an, and MiniMax, have already commenced adjusting their pricing strategies.

Zhipu has raised API prices three times this year, with its GLM Coding Plan package for developers witnessing an overall increase of at least 30%. Its flagship models, GLM-5-Turbo and GLM-4, have also seen price hikes of 20% and 10%, respectively.

When Yuezhia'an unveiled its 2.8-trillion-parameter Kimi K3 in mid-July, it raised the pricing of its new-generation API by 3 to 4 times compared to the previous version, with output prices directly reaching 100 yuan per million tokens.

Model training, inference, and infrastructure costs continue to escalate. Every large-scale model vendor is assessing whether the market share gained through low prices can translate into sustainable revenue.

Compared to other large-scale model vendors, DeepSeek's price adjustment confronts a more intricate reality.

DeepSeek faces a challenge that closed-source vendors do not: the weights of V4 Flash are publicly released under the MIT license, enabling third parties to self-host, modify, and commercially distribute the model. Customers can utilize DeepSeek's official API, rent from other service providers, or run it on their private infrastructure.

In other words, DeepSeek cannot price like a traditional software monopolist.

While its official service can charge a premium for convenience, optimized inference speed, high reliability, instant upgrades, and enterprise-level support, the open-source weights establish a natural pricing ceiling. If DeepSeek raises prices too aggressively, third-party hosting providers can directly undercut it using DeepSeek's own open-source model.

Despite these hurdles, DeepSeek has opted to raise prices, driven by deeper motivations likely linked to capital market expectations for its commercialization prowess.

As previously mentioned, DeepSeek has initiated preparations for going public, with a new round of financing underway. As capital investment surges, the criteria for evaluating DeepSeek's value are shifting.

In the past, the focus was on model capabilities, benchmark scores, user scale, and industry influence. However, in the face of massive computational investments, investors now prioritize sustainable commercial metrics.

For DeepSeek, low prices have aided in completing market education. The next step is to demonstrate that it can generate profits—and sustain them.