Claude Code with a local LLM as the worker: cheaper coding

Frontier models are expensive and subscription limits keep shrinking, so the author tries a split: Claude writes small task prompts, reviews the output against the spec and sends defects back, while a 27-billion-parameter Qwen model, quantized and running on llama.cpp with a 32K context, writes the code. The worker ran on a single rented 24 GB GPU at $0.57 an hour.

The test project was a five-module Python key-value store with 78 tests, 58 of them hidden from the worker. The two orchestrated runs passed 78 and 77. Three one-shot runs, given the whole spec in one prompt, passed 76, 74 and 71. Every one-shot run shipped the same bug (restarting a session rewrote its log), and so did both orchestrated runs until the review caught it. Orchestration took three to eight times as many worker turns but kept every prompt under 8K tokens. The GPU time for the whole session came to about forty cents.

The author’s conclusion is modest: the cheap model is good enough for small, well-specified tasks, the review step is where most defects were caught, and they would not trust the setup with something like a database. The caveats are stated plainly: only two or three runs per condition, a second review that was not fully blind, a precise spec that probably helped, Claude’s own token costs not counted, and no way to separate the effect of splitting the work from the effect of reviewing it.