What Is DeepSeek Context Caching?
DeepSeek context caching (also called disk caching) stores the computation for prompt prefixes and reuses it on matching requests, billing cached input at about 1/50 of the normal input price.
How prefix matching works
DeepSeek caches the computation results for prompts starting from the first token. When you send a new request, the system checks whether the beginning of your prompt matches a previously cached prefix. If it does, the cached computation is reused — and you're charged the cache-hit price instead of the full input price.
The critical detail: matching is prefix-based and starts from token 1. Change anything at the beginning of your prompt — even a single character — and the entire cache is invalidated.
The pricing difference
As of September 10, 2026, DeepSeek V4 Flash charges ¥2.00 per million input tokens for cache misses at peak (¥1.00 off-peak), but only ¥0.04 per million for cache hits (¥0.02 off-peak) — a 98% discount. For V4 Pro the ratio is roughly the same: ¥9.00 miss vs ¥0.30 hit at peak. Off-peak, both are halved again.
This makes caching the single biggest cost lever in DeepSeek's billing model. A stable system prompt plus reusable context is effectively free to send, while unique one-off prompts pay full input price.
How to exploit it
Keep system prompts absolutely fixed and placed at the front of every request. Put dynamic content — user queries, timestamps, results — after the reusable prefix. Version prompt changes deliberately and watch the hit-rate trend after each change.
The four usage CSV types make caching visible: input_cache_hit_tokens and input_cache_miss_tokens are exported per request, so you can measure your real hit rate and per-key cache economics from billing data alone.
Frequently Asked Questions
What is the difference between context caching and cache hit rate?
Context caching is the mechanism (storing and reusing prefix computation); cache hit rate is the metric (the share of input tokens billed at the reduced rate). Good prompt structure raises the hit rate, which is how you capture caching's savings.
How much does DeepSeek context caching save?
Cached input is billed at about 1/50 of the uncached input rate — a ~98% discount. On a high-hit-rate workload this can cut total input cost by more than half.
Does changing my system prompt break the cache?
Yes. DeepSeek caching is prefix-based from token 1, so any change at the start of the prompt invalidates the whole cache. Treat prompt changes like code releases and monitor the hit-rate trend afterward.
Dive deeper
Related terms
Recommended Tools We ARE USING
We are using these tools ourselves for Development / Deployment. Check out for more details.
Opencode Go
You will receive $5 when subscribed, which directly offsets the 1st month's Go fee.
Could be the cheapest option for calling DeepSeek V4 Flash model in the world.
Vultr
RReceive $300 to test out VVultr platform
Referred user must be active 30+ days and spend $10–$25
Railway
Deploy anything without the complexity
Connect your repo, Railway handles the rest.
腾讯云 Tencent Cloud
Cloud servers, cloud databases, COS, CDN, SMS and other cloud products are on special offer now.
AWS high-quality alternative, we use COS as a replacement for S3
硅基流动 SiliconFlow
¥16 universal voucher for all platforms, valid for 180 days from the date of receipt
A comprehensive product matrix supporting the full-process implementation of AI applications.
Warp
One of the best AI terminal tools in the world
Interestingly, we can directly use various top-tier models in Warp without leaving the terminal.