Another "Harness matters" post (codex cli > pi and opencode)

A user running Qwen 3.8 flash Next on 2x3090s and RAM notes that local LLMs, despite their speed, have not performed as well as GPT 5.2 or 5.6 Luna for serious work. They use a "3D Mario Game, multistage" demo to test LLMs and wonder if they are using pi.dev incorrectly or if an extension could bring it to the quality of Codex Cli.

This user's experience highlights a performance gap between local LLMs like Qwen 3.8 flash Next and proprietary models such as GPT 5.2 or 5.6 Luna, unlike benchmarks that might suggest otherwise.