Unexpectedly,DeepSeek’s API price war has only just begun.The company has changed its strategy and is now announcing that prices will be going up, possibly by a significant margin. Recently, DeepSeek has begun notifying API customers that it will substantially adjust overall API service pricing in the future, though the actual new rates and effective dates are not yet known.
This move also makes one wonder if DeepSeek’s own GPU resources are about to give out.
DeepSeek API Announces Significant Price Increase: Demand Surges After V4 Flash Price Cut, Official New Rates Not Yet Announced
According to user @gpuhell on X, DeepSeek has begun notifying API customers that it will raise overall API service prices, and the increase is expected to be “substantial”:
Currently DeepSeek Official Pricing PageThere is also a special note that plans are underway to adjust overall API pricing in the near future, with an expected significant increase. The specific plan will still depend on the official announcement, and developers are reminded to plan their usage based on actual consumption in advance.
DeepSeek used the unusual phrasing “significant increase” in this announcement, so DeepSeek API users really should be mentally prepared, and even start thinking about alternative plans in case of the price hike.
The current prices are: for DeepSeek-V4-Flash-0731, cache-hit input tokens cost 0.02 RMB per million, cache-miss input tokens cost 1 RMB, and output tokens cost 2 RMB; for DeepSeek-V4-Pro, they are 0.025 RMB, 3 RMB, and 6 RMB respectively.
| It appears you’ve provided only the instruction without the Traditional Chinese text to translate. Please provide the source text. | Cache hit input | Cache miss input | Output |
|---|---|---|---|
| DeepSeek-V4-Flash | 0.02 yuan | 1 yuan | 2 yuan |
| DeepSeek-V4-Pro | 0.025 yuan | 3 yuan | 6 yuan |
So why did DeepSeek just lower prices, only to suddenly announce a price increase?
At this point, the most likely cause is that V4 Flash’s low-price strategy has generated a surge in demand in a short period, putting pressure on server-side inference capacity.
A user mentioned on X that DeepSeek models have very slow inference speeds during certain periods. Given that DeepSeek Labs only has a limited compute scale of about 20,000 NVIDIA H100 GPUs, this is highly likely.
Following the earlier major price cuts, many users have likely integrated DeepSeek into AI agent tools. The lower the price, the easier it is for developers to scale up usage, but for model service providers, every output continuously consumes GPU inference resources.
When GPU computing power is insufficient, the only way to improve the situation is to raise API prices, which increases revenue while also reducing non-essential traffic, reserving capacity for users who truly need stable service.
For developers using the DeepSeek API, the most important thing next isn’t guessing how much prices will rise—it’s recommended to first figure out which parts of your current workflow are cache hits, which are general inputs, and what proportion of your monthly usage is output tokens. That way, you can respond more quickly when changes come.
Source: KOCPC Chinese