I built vix, an AI coding agent that uses a virtual filesystem to minify source code so the LLM always works in the minified space, and the concept of stem agents that allows for optimal cache hits in plan mode.
I used the same prompts and same model than claude code, and run both vix and claude code over 7 life like benchmarks, ran 5 times each. The results are that vix is 50% cheaper and 40% faster than claude code. I shared all the plans and the diffs in the repo so you can compare the planning and execution quality by yourself (similar in my opinion).
Comments URL: https://news.ycombinator.com/item?id=47680184
Points: 1
# Comments: 0