Luna Max — Workhorse!

Hello everyone! I realize I might be beating a dead horse here, but I wanted to offer my experiences as someone who’s been on a ChatGPT Plus plan for quite some time, and has used Sol-High and Luna-Max extensively.

These are just my own thoughts of course, and feel free to add commentary about your experiences with both GPT and GLM/Deepseek, since I’ve tried all of them! ^o^

Of course, we can’t talk about Sol on codex without talking about the usage limits. The usage limits for Sol are astounding; it eats tokens for breakfast and deletes them for dinner! It is very good at planning, but execution is sometimes over-engineered to the point where human-reading the code is impossible unless you’re very experienced with complex codebases or you just keep using AI for the entire codebase as a result.

What I’ve noticed about Luna (I only use MAX as it is a lower-intelligence model, and MAX doesn’t seem to burn quota any more than the lesser thinking levels) is that it can get jobs done fairly quickly, concisely, and efficiently. I’m comparing this to models like GLM 5.3 (slow currently unless using Flash), and Deepseek V4 (pro/flash; I have ~15 USD deposited into the website since I use it from time to time). The only cases where Luna struggles is sometimes with multi-layered plans as well as in weird random cases. I had an issue that when editing a text bubble, it had a minimal width/height instead of the regular width/height of the text bubble, and it took three prompts to get them to match..

The main reason to use Sol at this point is for projects which require ‘true’ reasoning (that’s what I call it, anyway). What do I mean by this? I mean that projects which have never been seen before by the training data of the model. Luna and Terra, along with GLM and Deepseek, fail spectacularly at these sorts of problems for whatever reason. My example is making a script for a game to automate something, and the script has never existed before, with the video game also not in the mainstream, so it’s unlikely in the training data. The only model which can accurately make a plan, execute on it, and even edit the existing script is Sol and Sol only.

(not sure about GPT-6.. not paying API pricing and it’s not on Codex plan yet)

What are your thoughts? Maybe I’m using GLM or Deepseek wrong, and there are prompts that enhance their workflow? Or maybe MCP that I can configure for these models for them to be smarter? Let me know!