Octobench finds a five-task harness swing

Octobench put GLM-5.2 through the same 25 fresh pull-request tasks and found a five-task harness swing: Octomind solved 24 for $63.43, while OpenCode solved 19 for $129.54 on the same endpoint and prices, according to the benchmark’s pinned results. The vendor-authored, single-run comparison is too small to crown a winner, but it sharpens the argument that agent harnesses—not model weights alone—shape coding outcomes: builders should A/B context retrieval and completion guards before paying to swap models.