AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test
On real change requests the best model got 26.4 percent of jobs fully right, and GLM-5.2 burned nearly six times as many tokens doing significantly worse, which the meter records as a good day.