Benchmark across Claude, OpenCode, Hermes, and other coding agents on 10 SWE-bench tasks

Carlo Capocasa published a 10-task GLM benchmark comparing Claude, OpenCode, Pi, Zcode, Hermes, and 3code on representative SWE-bench verified tasks. 3code solved 9 of 10 tasks using 5 million tokens, while Pi solved 6 of 10 using fewer tokens than other runners-up. Capocasa noted harness performance varies with token efficiency and task completion rates, cautioning that users should validate results against their own heuristics.