DeepSeek V4 Pro is cheap only when the harness preserves the prefix

[bobleer]

DeepSeek V4 Pro separates cache-hit and cache-miss input pricing. Whether a long coding run reaches the cheap path is largely a harness question.

The DeepSeek cache guide says a later request must fully match a persisted prefix unit. In an agent loop, a timestamp, reordered tool schema, or nondeterministic prompt assembly can turn a growing context into cache misses.

I contribute to BitFun. We made prompt assembly byte-stable across turns, then measured a 98.67% average KV-cache hit rate over one SWE-Bench-Pro run with DeepSeek V4 Pro. This is one run average, not a promise that every task stays at 99%.

This matters when an agent repeatedly sends repository context while it reads code, edits files, runs tests, and revisits failures. Model pricing sets the unit cost; prefix stability determines how often the lower-cost path is actually used.

Evidence:

• DeepSeek pricing: https://api-docs.deepseek.com/quick_start/pricing
• DeepSeek cache matching rules: https://api-docs.deepseek.com/guides/kv_cache
• BitFun measurement and scope: https://github.com/GCWing/BitFun#agent-core-metrics
• Source: https://github.com/GCWing/BitFun

BitFun is an open-source desktop agent for real repository workflows. If this measurement and byte-stability work are useful, give the repo a star. I would also like to compare real hit/miss traces from other harnesses.