Summary of Key Points
DeepSeek plans to launch the official V4 version of its large-scale model API in mid-July, introducing a "peak-valley pricing" mechanism. The main working hours in China (9:00-12:00, 14:00-18:00) will be considered peak periods, during which prices will double compared to normal times; outside of these peak periods, the prices will remain the same as those in the previous preview version (with no overall increase). This adjustment has sparked heated discussions among developers. Some criticize it as unfair to Chinese users, while others understand the cost logic behind it. The industry sees this as a sign of maturity in AI cloud services, suggesting that other domestic large-scale models may also adopt flexible pricing in the future. Developers will need to manage their costs through task scheduling, and there is general anticipation for improvements in the capabilities of the V4 official version.
Detailed Explanation
1. What exactly is peak-valley pricing?
In simple terms, DeepSeek divides the day into "busy" and "idle" times:
- Peak periods: From 9:00-12:00 and 14:00-18:00 Beijing time (when most people are working), the API prices will double.
- Non-peak periods: The rest of the time (e.g., at night or early in the morning) will have the same low prices as in the previous preview version.
For example, using the V4 Pro model, processing data without caching will cost 6 yuan per million tokens during peak times and 3 yuan per million tokens during non-peak times. This means that using AI during working hours will be twice as expensive, but it will be just as cheap as before at night.
2. Developers are divided: some complain about bias towards overseas users, others think it's acceptable
Developers have different reactions:
- Complaints: Most Chinese users are working during peak periods, so they will have to pay more. However, since users in Europe and America are in non-peak times (e.g., evening in their time zone is morning in China), this seems like a benefit for overseas users, which some consider unfair to Chinese developers.
- Supporters: DeepSeek's current pricing is already lower than the industry average (for example, much cheaper than GPT). Even with the price doubling during peak periods, it remains competitive. Developers are more concerned about whether the V4 official version will have improved capabilities, such as enhanced programming or multi-modal features. If the capabilities are strong enough, they are willing to pay a higher price.
3. Is this a sign of AI industry maturity?
Peak-valley pricing is not arbitrary but reflects the need for more efficient resource allocation in the AI sector:
- AI models require significant computing power (e.g., GPUs), which is in high demand during peak periods and wasted during off-peak times. This pricing helps to allocate resources more reasonably, improving service stability.
- Industry experts predict that other domestic large-scale models (such as Wenxin Yiyan and Tongyi Qianwen) will likely follow suit with flexible pricing, similar to how electricity costs vary by time of day.
4. Developers need to become "time management masters" to save money
Peak-valley pricing forces developers to adjust their work habits:
- Tasks that don't require immediate attention, such as batch data processing or model training, can be scheduled for off-peak times to reduce costs by half.
- It's also important to use caching: storing frequently used input data for quick access later, which can significantly lower prices (e.g., 0.05 yuan per million tokens for cached inputs during peak periods, compared to 6 yuan for uncached ones).
- Task scheduling capabilities (automatically scheduling tasks during off-peak times) will become essential for cost control.
5. The biggest question: Will the V4 official version have improved capabilities?
Developers are more interested in the actual improvements of the V4 official version, such as enhanced "post-training" features, multi-modal capabilities, improved programming abilities, and advanced agent functions.
- Many users believe that if the capabilities are significantly better, higher prices during peak periods are acceptable, as the value of AI tools lies in increasing efficiency.
In summary
DeepSeek's peak-valley pricing marks a shift from "extensive" to "precision-oriented" pricing in the AI industry. It optimizes resource allocation and reminds developers that choosing the right AI model is just as important as choosing the right time to use it. However, whether users will accept this pricing depends on whether the V4 official version truly offers improved capabilities. After all, the ultimate test of any AI tool is its effectiveness in helping people improve their efficiency.