I'd imagine it's just a lot less token efficient than having it write the code in a high level language.