How a 99.82% Cache Hit Rate Is Achieved — DeepSeek Prefix Cache Engineering Explained

Why 99.82% Is Not an Accident

A published usage log from an AI coding session using DeepSeek Reasonix showed a 99.82% prefix cache hit rate on 435 million input tokens — reducing the effective cost of that session from roughly $61 to approximately $12. That number is not a product of luck. It is the result of specific, intentional engineering decisions in how the agent constructs and maintains its prompts.

To understand what makes that figure achievable — and why most coding agents land at 20–60% — you need to understand how DeepSeek's prefix cache works.

How DeepSeek's Prefix Cache Works

DeepSeek's API maintains a server-side KV cache of recent prompt prefixes. When a new request arrives, the API checks whether the beginning of the prompt (the "prefix") matches a cached computation. If it does, the model does not recompute attention over the cached portion — it picks up from the cache boundary and processes only the new tokens.

The cache is byte-stable: the prefix must be exactly identical, byte for byte, to register a hit. A single character difference — a whitespace change, a timestamp update, a reformatted tool call — invalidates the cache for everything after that point. This is why most agents fail to sustain high hit rates: their prompt construction is not stable across turns.

The Four Failure Modes That Destroy Cache Hit Rates

1. Dynamic system prompts. Including timestamps, session IDs, user names, or any value that changes between turns in the system prompt guarantees a cache miss on every request. Many out-of-the-box agent frameworks insert dynamic values into system prompts for tracing or personalization without realizing the cache cost.

2. Inconsistent tool call formatting. JSON serialization of tool call arguments can produce different byte sequences for the same logical content depending on key ordering, whitespace, or number formatting. If your agent library re-serializes tool results between turns, the prompt changes even when the data does not.

3. File context reinsertion with modification timestamps. Agents that include file metadata (last modified time, git hash, inode) alongside file content in the context will invalidate the cache every time a file is saved — including mid-session saves triggered by the agent itself.

4. Conversation history truncation strategies. Agents that trim conversation history by dropping the oldest messages from the middle of the array break the prefix structure. The system prompt and early turns are shared across sessions — truncating from anywhere other than the tail destroys cache continuity.

What Reasonix Does Differently

Reasonix's "cache-first loop" addresses all four failure modes:

The system prompt is static and deterministic — no timestamps, session IDs, or user-specific content. File context is inserted as raw content only, with no metadata. Tool call results are serialized through a canonical formatter that guarantees byte-identical output for identical logical content. Conversation history is always trimmed from the tail only, preserving the prefix.

The compound effect of these four decisions is a prompt prefix that remains byte-identical across hundreds of turns in a session, which is why the cache hit rate approaches 100% rather than the 20–60% range typical of general-purpose agents.

The Cost Math at Different Hit Rates

The cache hit rate is the single most important variable in DeepSeek API costs for long coding sessions. A 40-percentage-point improvement in hit rate (from 60% to ~100%) reduces costs by 20× on input tokens — dwarfing the impact of any model tier choice.

Applying This to Your Own Agent

If you are building on top of DeepSeek's API directly, the most impactful change you can make is auditing your system prompt for dynamic content and ensuring your serialization layer produces stable byte sequences. These changes typically take under a day to implement and have an immediate impact on your API bill.

For developers using existing agent frameworks, check whether your framework exposes cache hit rates in its usage metadata. If you are seeing consistent cache misses on requests where the content should be identical, the four failure modes above are the most likely causes.

Use the AI Cost Estimator to model how improving your cache hit rate would affect your monthly DeepSeek API spend.