What Is DeepSeek Cache Hit Rate?

DeepSeek cache hit rate is the share of your input tokens that are billed at the reduced cache-hit price instead of the full input price — the higher the better, since cached input costs as little as 1/50 of the uncached rate.

How it's measured

DeepSeek's usage exports include two input-token rows per request: input_cache_hit_tokens and input_cache_miss_tokens. The cache hit rate is simply cache hits divided by total input tokens (hits + misses).

A hit rate of 50% means half of your input tokens rode the cached lane. Because DeepSeek bills cache hits at roughly 1/50 of the uncached rate, that single number has an outsized effect on your bill: raising it from 20% to 60% can cut input costs by more than half.

What a good rate looks like

Real-world DeepSeek workloads routinely sit between 40% and 96%. On OpenCode Go, for example, DeepSeek V4 Flash holds a 96% real-world cache-hit rate, which is why its effective cost is far below the sticker price.

As a practical band: below 20% usually means your stable prefix is too short or variable content appears too early in the prompt; 20–40% is workable but has room; 40–60% is strong; 60%+ is excellent for repetitive workflows like customer support and templated extraction.

How to raise it

DeepSeek uses prefix matching: it caches computation starting from the first token, so anything that changes at the front of your prompt invalidates the cache. Keep system prompts fixed, place dynamic content last, and version prompt changes deliberately.

The easiest way to see your real rate is to upload your billing CSVs to a usage dashboard — the charts will show your daily cache hit rate and per-key hits vs misses at a glance.

Frequently Asked Questions

Dive deeper

Related terms