Same Model Benchmark: Agent A (45/50) vs Agent B (43/50) with Similar Cost

donk8r · reddit · 2026-08-22

A benchmark test compared two open-source coding agents (Octomind vs. a competitor) using identical settings (deepseek-v4-flash, same prompts, no web access) across 50 tasks. Results show Octomind scored 45/50 vs. 43/50, with judge scores of 88.6 vs. 85.6, and near-identical costs ($1.59 vs. $1.53). Although Octomind had a slower mean time, its median was faster (5.5m vs 7.0m), winning in 31 of 50 cases. The mean time discrepancy was caused by a single extreme failure case, revealing a bug in Octomind's no-progress detection.

Original post →

More from coding & agent

coding & agent channel →