In this post, Rui Carmo shares his experiences with the regressions in the GPT-5 models, particularly noting how GPT-5.5 produces less reliable code fixes compared to its predecessor, GPT-5.3 Codex. He reflects on a pattern where newer models become broader in functionality but sacrifice depth and precision in code analysis. Carmo recommends using older models for code audits while utilizing newer ones for tasks requiring broader capabilities, emphasizing the importance of swappable model systems in development.