LEVI: Open-source AlphaEvolve runs 35x cheaper on Claude Code

A developer known as Longjumping-Music638 has introduced LEVI, an open-source system that emulates the AlphaEvolve approach but at a fraction of the cost. After six months of work, LEVI claims to be up to 35x cheaper than existing open-source frameworks while delivering superior performance. It runs on Claude Code or Codex, and the developer primarily used it with a QWEN-30B model. The system targets two main domains: code optimization and prompt optimization, with a clear design philosophy that smaller models with the right search architecture can replace larger, more expensive ones.

LEVI achieves its efficiency through three key mechanisms: investing in solution diversity from the start to avoid convergence to the same solution, using smarter routing to decide when to use larger models (most mutations don't require Claude Opus), and building proxy subsets for prompt optimization to approximate results without expensive rollouts. In tests on systems problems like MoE scheduling and database transaction scheduling, LEVI outperformed existing frameworks while using up to 7x fewer resources. For prompt optimization on benchmarks like IFBench and HotSpotQA, it matched or exceeded GEPA with less than half the rollouts. The project is available on GitHub with a detailed technical blog.