Open Source

DeepSeek V4 Flash harness showdown: speed varies 4x, code quality unchanged

Claude Code took nearly 4x longer than Pi for identical diffs — the harness matters more than the model.

Deep Dive

A developer benchmarked DeepSeek V4 Flash (running on vLLM at ~180 tok/s) across three popular AI coding harnesses—Claude Code, OpenCode, and Pi—using the same antigenic workload on a large codebase. The results were clear: code diff quality was essentially identical across all three, but the path to get there differed dramatically. Claude Code was the slowest, taking nearly four times longer than the fastest harness (Pi), largely because it explored the codebase excessively, making more tool calls and burning tokens on unnecessary traversal.

OpenCode took a delegation-driven approach, while Pi focused on reasoning before acting. The system prompt and tool structure of each harness drove these efficiency differences, not the underlying model. The developer later expanded the test to include NanoCoder, reinforcing the pattern: harness design impacts speed and cost far more than output quality. Full raw data and charts are available on the author's site, inviting the community to dig deeper.

Key Points
  • Claude Code was the slowest harness, taking ~4x longer than Pi for the same diffs.
  • Code diff quality was identical across Claude Code, OpenCode, Pi, and NanoCoder.
  • Harness efficiency varied widely due to differing tool-call structures and system prompts, not model performance.

Why It Matters

Developers can save significant time and compute cost by choosing the right harness, without sacrificing code quality.

📬 Get the top 10 AI stories daily