DeepSeek V4 Pro 0813 has made cost comparisons fun again. The model is capable, the API is cheap, and the first wave of coding demos looks good.
But model price is not the same thing as agent task price. A harness can quietly erase much of the advantage before the model gets to do useful work.
The one-byte problem
A coding agent resends a large prefix on every turn: system instructions, tool schemas, repository context, and conversation history. Prefix caching only helps while that prefix stays identical. Add a timestamp, reorder the tools, or serialize one object differently and the cache can miss from that byte onward.
That matters more on real repository work than on a one-shot demo. Search → patch → test → read the failure → repeat can run for dozens of turns. If every turn rebills the same prefix, a cheap model starts behaving like an expensive workflow.
What we measured in BitFun
We build BitFun, an MIT-licensed desktop agent workbench with a Rust runtime. Prompt assembly is kept byte-stable across turns. In one SWE-Bench-Pro run using DeepSeek V4 Pro, the average KV-cache hit rate was 98.67%. This is a single-run engineering signal, not a universal benchmark claim, but it tells us the optimization is doing its job.
The other half is persistence. Long repository tasks rarely finish in one clean response, so BitFun's Goal mode persists the objective across turns and automatically continues while it remains active. The runtime can schedule up to 100 continuation turns for one objective, with bounded retry backoff when a continuation submission fails. It does not promise that every task magically succeeds; it is designed so the agent does not confuse “I wrote a plausible answer” with “the repository task is actually done.”
Why V4 Pro and the harness should be evaluated together
The same model can look very different in two coding tools because the model only supplies part of the system. Tool selection, error recovery, context packaging, repository search, and stopping policy all live in the harness.
BitFun already includes DeepSeek V4 Pro in its model catalog, but support is the boring part. The interesting question is whether the surrounding runtime preserves DeepSeek's cost advantage through a long search/patch/test loop. That is the problem the 98.67% measurement answers.
BitFun runs on macOS, Windows, and Linux, and the source plus installers are here: github.com/GCWing/BitFun. If this is the kind of open agent workbench you want to see grow, a star helps more builders find it.