Benzi benchmarks code-fixing harnesses against Claude Code and DeepSeek on lines read, time, and cost

Show HN: Benzi – A Code Intillegence/Harness Beating Claude Code and CodeGraph

Benzi tested four setups on 24 real GitHub issues: Benzi with Sonnet or DeepSeek, Claude Code, and the DeepSeek native harness. The headline metric is source lines read per fix—Benzi + Sonnet read 9,125 lines total, Claude Code read 20,704, and the DeepSeek harness read 43,598. Cost-wise, Benzi + Sonnet spent $17.96 for all 24 bugs vs. $39.54 for Claude Code; Benzi + DeepSeek cost just $2.66. On SWE-bench Verified, Benzi resolved 78.2% of 500 issues at under 10¢ per fix. The post doesn't explain how Benzi's code intelligence achieves the lower read counts, and it doesn't break down latency details.