Choosing a coding agent involves more than choosing its model. In a September 13 analysis, developer Geoqiao compares Pi, Codex, Claude Code and other agent harnesses: the software that manages the model’s tools, context and repeated attempts to complete work.
The comparison brings together published evaluations and the author’s own project test. Some hold the model fixed while changing the harness. Others change the model or reasoning effort too, so those results describe the whole configuration rather than isolating the software around it.
The findings favor different choices for different jobs. Pi and Codex perform well across several of the cited comparisons, while Claude Code remains a candidate for native Claude workflows. In the author’s blog-theme exercise, completion is judged against 20 requirements; the cost figures are estimated API-equivalent costs for the whole development run.
The separate benchmarks use different tasks, repetitions and cost denominators. A price per successful task cannot be placed directly beside a price averaged across every attempt.