跪拜 Guibai
← All articles
DeepSeek

DeepSeek Slashes API Prices Permanently to One-Quarter of Original

By 飞哥数智谈 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A permanent price floor reset at one-quarter of the original rate, combined with near-total KV cache coverage, changes the unit economics of building on DeepSeek's API. Teams running high-volume inference workloads can expect daily costs to fall by nearly an order of magnitude compared to the pre-discount era, without the rate-limiting friction common on competing platforms.

Summary

DeepSeek Pro's API pricing will settle at one-quarter of the original rate once the current 75% promotional discount expires on May 31, 2026. The move turns what looked like a temporary sale into a permanent structural price cut. A daily workload of 5 million tokens—split 70% input and 30% output—drops from roughly 78 yuan to 19.5 yuan before caching, and to about 9.3 yuan when KV cache hits reach 98%.

KV caching is the quiet multiplier here. Cache hits on DeepSeek's API have been observed at 98% without any special tooling, and the cost per million cached input tokens is just 0.025 yuan. That makes sustained, high-volume inference dramatically cheaper than the headline input price suggests.

Other model providers still charge more than DeepSeek's original rates and frequently impose rate limits, which makes the combination of low base pricing and effective caching a direct challenge to the rest of the market.

Takeaways
DeepSeek Pro's API price will permanently settle at one-quarter of the original rate after the 75% promotional discount ends on May 31, 2026.
Input tokens drop from 12 yuan to 3 yuan per million; output tokens drop from 24 yuan to 6 yuan per million.
KV cache hits reached 98% in one user's testing without specialized tooling, and cached input costs only 0.025 yuan per million tokens.
A 5-million-token daily workload went from 78 yuan (pre-cut, no cache) to 9.3 yuan (post-cut, 98% cache hit rate).
Competing model providers still charge more than DeepSeek's original prices and frequently impose rate limits.
Conclusions

Framing a permanent price cut as the end of a limited-time discount is a clever marketing tactic that reframes a price war as a customer reward.

KV caching is the underappreciated lever in LLM API costs; near-100% cache hit rates make the effective input cost approach zero, which matters more than the headline token price for many workloads.

DeepSeek's pricing trajectory suggests the company is willing to run inference at near-cost to capture developer mindshare and usage volume, a strategy that pressures competitors who rely on API margins.

Concepts & terms
KV Cache
A key-value cache that stores the computed representations of previous input tokens so that repeated or overlapping prompts don't require recomputation, drastically reducing latency and cost for conversational or iterative workloads.
From the discussion

The discussion is thin. One comment quotes a planned price hike, directly contradicting the article's claim of a permanent cut. A reply objects to the article's timeliness, noting the news is from May.

A planned near-term price increase contradicts the article's framing of a permanent price cut.
The article is criticized for being outdated, republishing May news as if it were current.
Featured comments
ovenslove 4 likes

We plan to raise the overall pricing of the DeepSeek API service in the near future, with a significant increase expected. Please plan your usage accordingly. The specific plan is subject to the official notice.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗