Author｜Zhang Qiming

Editor｜Wang Yimo

// Agents Drive Up Computational Demand, Industry Moves Toward Cost Recovery

On August 6, DeepSeek revealed its intention to substantially raise the prices of its DeepSeek API services in the coming days. This move, combined with its existing peak-valley pricing system—which doubles fees during peak usage times—clearly indicates a trend toward higher costs for AI interfaces.

The surge in Agent usage has led to increased computational demands, prompting the domestic large-scale model industry to move away from price wars.

The price adjustment was communicated through an official notice to developers on the DeepSeek open platform, urging them to strategize their API usage and recharge schedules wisely. Full details regarding the extent of the price hike and its effective date are still pending official confirmation.

Previously, DeepSeek pioneered the implementation of peak-valley pricing within the industry. During weekday peak hours (9:00-12:00 and 14:00-18:00), all API call fees are subject to a twofold increase.

During off-peak hours, the cost for cached input hits with V4-Flash is merely 0.02 yuan per million Tokens, with non-cached inputs costing 1 yuan and outputs 2 yuan. For V4-Pro, cached input hits are priced at 0.025 yuan per million Tokens, with non-cached inputs at 3 yuan and outputs at 6 yuan.

The latest price adjustment will further elevate these base rates, in addition to the peak-valley pricing mechanism. Once implemented, the combined impact of peak pricing and base rate increases will lead to a notable surge in overall API call costs.

Whether the prices of consumer-facing conversational membership products will also rise remains to be officially confirmed.

As a prominent domestic provider of large-scale models, DeepSeek's V4-Flash and V4-Pro models are renowned for their cost-effectiveness and are extensively utilized by AI startups, government and corporate systems, and Agent development projects. The cost structures of many small and medium-sized developers' products heavily rely on this affordable API system.

The combination of peak-valley pricing doubling and the impending comprehensive price increases is swiftly closing the window of opportunity for low-cost Token experimentation.

The price hikes reflect a fundamental shift in the industry's supply-demand dynamics.

The widespread adoption of Agent intelligent systems has introduced a machine-driven calling pattern, where complex tasks generate multiple rounds of thinking, tool calls, and iterative error corrections. This results in an exponential increase in Token consumption compared to traditional conversations, driving up overall inference computational power usage.

The global supply of high-end GPUs and HBM memory remains constrained, with hardware depreciation, electricity, and data center maintenance costs representing fixed expenses. As call volumes rise, the pressure on vendors' profitability intensifies, rendering the low-cost subsidy model unsustainable.

DeepSeek's move is not an isolated incident; in 2026, a wave of API price adjustments has swept across China.

Zhipu AI has increased its interface pricing multiple times throughout the year, with some interfaces in Tencent Cloud's Hunyuan series experiencing price hikes exceeding 460%. Alibaba Cloud and Baidu Intelligent Cloud have also raised prices for their AI computational power-related products and services. Leading overseas model vendors have similarly increased their service prices, making price hikes a prevalent trend across the industry.

Over the past two years, the industry has predominantly employed a "price war for market share" strategy, relying on capital subsidies to lower interface prices and attract developers.

However, the industry's focus has now shifted toward model capability enhancement and Agent business expansion, necessitating increased computational power investment. Commercial viability has taken precedence over merely pursuing user scale, making it inevitable to pass on costs to downstream developers.

At this industry inflection point, market players will experience divergent outcomes.

Large enterprise clients, with their long-term computational power reserves and multi-model routing strategies, can mitigate the impact of single-model price hikes and continue to enjoy stable services. The hardest hit will be small and medium-sized startup teams that heavily rely on third-party APIs and lack their own computational power reserves.

With their original business models built on low-cost Tokens, these developers face rapidly diminishing gross margins and even the risk of "losses with every call" following cost increases. Developers will need to revamp their cost calculation systems through strategies such as cache reuse, off-peak scheduling, prompt optimization, and multi-model hybrid routing.

At the industry level, price hikes will also reshape competitive opportunities. Startup projects that focus solely on the application layer and rely entirely on external large model APIs will encounter higher survival thresholds. In contrast, enterprises with vertical scene expertise, private deployment capabilities, and self-developed lightweight models will see their competitive advantages further accentuated.

The criteria for capital evaluation of AI projects have also evolved, shifting from a sole focus on user growth to emphasizing the viability of the unit economic model.

Of course, risks persist. If API costs rise too rapidly, it may stifle AI application innovations and delay product launches for small and medium-sized teams.

Should consumer-facing membership services subsequently raise their prices, it could also dampen ordinary users' willingness to engage.

DeepSeek's price hike announcement is not merely a commercial maneuver by a single company but a reflection of broader industry trends.

With the advent of the Agent era, computational power can no longer be consumed indefinitely without cost. The entire AI industry is transitioning from a phase of rapid expansion fueled by capital subsidies to a new stage of commercialization where costs and values are aligned.

Going forward, the competition among models will no longer revolve around who offers the lowest Token prices but rather who can deliver tangible business value to customers that justifies the costs.