Benzi - Harness/AI agent beats big players on benchmarks while reading less source code

Benzi, an AI agent, has demonstrated superior performance in benchmarks compared to major players like Claude Code, DeepSeek, and OpenCode. It achieves this by processing significantly less source code; Benzi Sonnet reads 9,125 lines, while Claude Code Sonnet reads 20,704, and DeepSeek's harness reads 43,598. OpenCode, with over 65,000 lines of code, was disqualified due to repeated failures. Benzi completes tasks faster and more cost-effectively, with a live demo available for testing its code intelligence.