Compiler backed agent beats Claude Code, OpenCode & other major agents and harnesses (benchmarks linked)

Benzi Sonnet, a compiler-backed agent, has outperformed major agents like Claude Code Sonnet and OpenCode in benchmarks. It achieves tasks faster and cheaper by reading significantly less source code—9,125 lines compared to Claude Code Sonnet's 20,704 lines and DeepSeek's harness's 43,598 lines. OpenCode, which reads over 65,000 lines, was disqualified due to repeated failures. Benzi Sonnet achieved 78.2% SWE-bench Verified, demonstrating the efficiency of indexing a repository over reading raw source code.