Deepseek officially announced a price reduction! Cache hit price is now 0.02 yuan, miss price is 1 yuan, a 50-fold price difference.

DeepSeek has once again adjusted its API pricing strategy. This time, it's not a simple across-the-board price reduction, but rather a focus on "cache hits." At the same time, it introduces differentiated billing for peak and off-peak hours, further targeting high-frequency calls and long-context scenarios.

According to the DeepSeek Open Platform announcement, the Flash series will implement new pricing starting at 12:00 PM Beijing time on September 10, 2026. During off-peak hours, the price per cache hit will drop to 0.02 yuan/thousand tokens, while the price per cache miss will be 1 yuan, a 50-fold difference; the price per output will be 4 yuan. During peak hours (Monday to Friday, 9:00-12:00 and 14:00-18:00 Beijing time), all prices will be double those during off-peak hours.

This means that for applications that can repeatedly reuse the same context, API costs can decrease significantly. Simply put, a "cache hit" means that content already processed by the model doesn't need to be fully recalculated in subsequent requests; the system can directly reuse the results. For scenarios with long prompts, fixed system commands, or continuous dialogues, this content often constitutes the majority of the input token; therefore, a higher cache hit rate results in a more significant cost reduction.

New pricing structure: letting prices follow usage patterns

The most significant change in this adjustment is that API pricing is no longer determined solely by the number of calls, but is now linked to both the call time and whether the content is repetitive .

For developers, this means that if most of the content can be cached, the cost of calling 1 million tokens each time could be far lower than re-entering them every time. Applications such as RAG, code assistants, customer service robots, and agents, which often have a large number of repetitive system commands, knowledge bases, and historical contexts, are more likely to benefit from this mechanism.

This will also change how developers optimize costs. In the past, the main approach was to choose cheaper models and reduce token consumption; now, they can further improve cache hit rates by fixing contexts and splitting dynamic content. In other words, API cost management is shifting from "using fewer tokens" to "reusing tokens as much as possible."

Behind the price cuts: both to attract users and to adjust computing power.

From the platform's perspective, caching and time-limited discounts are not just about giving users a "discount"; there is also a clear logic behind computing power scheduling.

Cache hits mean that some inputs do not need to be calculated repeatedly, which can reduce the computational cost of repeatedly processing the same requests; while off-peak discounts encourage users to avoid peak hours by shifting some non-real-time tasks away from weekday peaks. The combination of the two essentially improves the utilization efficiency of limited computing resources.

For DeepSeek, this model has another layer of meaning: it prioritizes offering discounts to users who can generate stable, high-frequency calls, rather than indiscriminately lowering prices for all requests. Applications such as RAGs, code assistants, customer service, and agents, once built on top of the API, typically experience migration costs that increase with call volume and business depth; therefore, these users have higher long-term value.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.