跪拜 Guibai
← All articles
Frontend · Backend · Artificial Intelligence

DeepSeek API Prices Jump 11× Overnight, Ending the Subsidized Era

By 程序员_小雨 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A generation of small tools and agent workflows was built on DeepSeek's unsustainably low API prices. Those cost assumptions are now broken, forcing developers to re-architect scheduling, model selection, and caching or watch their bills triple.

Summary

DeepSeek's latest API pricing model has gone live, instantly reshaping the cost landscape for developers who built on its famously cheap tokens. The flagship V4-Pro model now charges 0.3 yuan per million tokens for cache hits during peak hours, up from 0.025 yuan — an 11-fold increase. Standard input tripled to 9 yuan, and output jumped 3.5× to 27 yuan. Off-peak usage, defined as evenings and weekends, is priced at half the peak rate.

The price hike ends a period of aggressive subsidies that saw DeepSeek slash prices by 75% just three months ago. Surging demand from agent-based workloads and the real cost of running a trillion-parameter MoE architecture made the old pricing unsustainable. The move aligns DeepSeek with competitors like Zhipu and Tencent's Hunyuan, which had already raised their rates.

For small teams and solo developers, the immediate impact is a cost model that no longer works. Workloads that once cost a few dozen yuan a month can now run two to three times higher during business hours. The practical advice is to batch non-urgent tasks into off-peak windows, downgrade simple jobs to the cheaper V4-Flash model, and aggressively cache repeated inputs.

Takeaways
Cache-hit input for V4-Pro rose from 0.025 to 0.3 yuan per million tokens, an 1,100% increase.
Standard input tripled to 9 yuan and output rose 3.5× to 27 yuan per million tokens during peak hours.
Peak hours are weekdays 9:00–12:00 and 14:00–18:00; all other times are off-peak and priced at 50%.
V4-Flash output costs 9 yuan per million tokens at peak, two-thirds cheaper than V4-Pro.
The price change applies only to the API; web and app chat users are unaffected.
Competitors Zhipu, Hunyuan, and Wenxin had already raised prices earlier in the year.
Conclusions

The 1,100% jump on cache hits is the real shock — it punishes the very optimization pattern DeepSeek previously encouraged.

DeepSeek's pricing now structurally rewards off-peak batch processing, which will shift how small teams schedule non-real-time workloads.

The end of subsidized pricing doesn't just raise costs; it forces a segmentation of tasks by model tier that many developers previously ignored because everything was cheap enough to run on the best model.

Concepts & terms
MoE (Mixture of Experts)
A model architecture where only a subset of parameters (experts) is activated per input token, allowing trillion-parameter models to run with lower per-token compute than a dense model of equivalent size.
Cache-hit input pricing
A discounted rate applied when an input prompt matches a previously cached computation, letting the provider skip re-processing and pass savings to the user. DeepSeek's cache-hit discount was the deepest in the market before this hike.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗