Dev spends $36 to process 800M tokens using DeepSeek agent on real C# codebase

A developer ran an informal benchmark by deploying DeepSeek Harness paired with DeepSeek V4 Pro on LovelaceSharp, a real C# arbitrary-precision math project, over several days in late August and early September. The agent produced 45 Git commits across 193 changed files during the observed development window, processing roughly 613 million tokens at a cost of $28.35. A near-99% cache-hit rate on input tokens kept costs low despite the large token volume. The experiment was designed to test sustained, architecture-aware coding work rather than the isolated single-issue fixes typical of standard benchmarks like SWE-bench. The author argues the meaningful measure is the actual change to the repository, not the token count itself.