New Compiler based agent cuts costs by 2x and improves code intelligence (78.2% SWE-bench Verified @ 0.1¢)

A new compiler-based agent, Benzi, has been introduced, demonstrating significant improvements in code intelligence and cost efficiency. It achieved a 78.2% score on SWE-bench Verified and reportedly cuts costs by 2x. The agent also reads substantially less source code compared to other models like Claude Code, DeepSeek Harness, and OpenCode, leading to faster and cheaper bug fixing, as detailed in benchmark results.

This report details a new compiler-based agent that reads 2x less source code than Claude Code and 4x less than DeepSeek Harness, unlike previous agents.